Skip to content
anythingai.space
02Model Landscape·131 min read

Model Taxonomy

This chapter provides a comprehensive guide to the model landscape as of November 6, 2026, covering model families, capabilities, and selection criteria for production systems.

Last verified: November 6, 2026. The model landscape evolves rapidly. Always cross-check with provider pricing pages and release notes.

November 2026 headline (October 1 to November 2): October filled the gaps September left, and most of the new risk sits in billing rules rather than benchmarks. Anthropic shipped Claude Haiku 5.5 (October 7, claude-haiku-5-5) at $0.10/$0.50 per 1M, GPT-6 Luna's list price, but any request whose prompt passes 100,000 tokens pays 5x on every token, cache reads included: the first exception to "Claude 4.6 and later are flat to 1M". It also cut Sonnet 5.5 cache reads to $0.10 (0.05x), took Opus 5.5 fast mode out of research preview (October 15; Opus 4.8 fast mode is removed November 30), launched Claude Models on Google Cloud (Anthropic managed) as a sixth platform, and shipped API version 2026-10-21, under which temperature, top_p, and top_k return a 400 at any value on every model; the new SDK majors send it by default. OpenAI shipped GPT-6.1 Terra (October 26, $1/$5), which brings back three-tier Luna/Terra/Sol routing, the trusted-access GPT-6 Rosalind (October 30, $15/$75), 12-hour prompt caching (October 20, writes at 2.5x input), and the Decisions API (GA October 22). It also gave about three months' notice that the remaining GPT-4o, GPT-4o-mini, GPT-4.1, and GPT-4.1-mini snapshots, and fine-tunes of them, shut down January 14, 2027. Google made Gemini 4 Argon GA (October 13, gemini-4-argon) at $2/$10 through November 30 and $4/$20 from December 1, with implicit caching only; announced Gemini 4 Carbon (October 28) for a subset of Fairwind cyber defenders, with no model ID; and rerouted gemini-3.7-flash to 3.8 Flash and gemini-3.5-flash to 3.6 Flash (October 8, posted in its changelog), so a pinned model string kept working while the model behind it changed. Meta shipped Muse Wonder (October 12, muse-wonder-1.0), a closed flagship at $4/$20 per 1M on the Meta Model API, Opus 5.5's list price; the stronger Muse Wonder Full is sold only through Muse Code subscriptions, with no API. There is still no Grok 5. Open weights got a new license gate: Mistral Large 4 (weights October 29) requires a separate commercial license for any Model-as-a-Service business, at any revenue, and Meta Muse Spark Open (October 15) gates MaaS above US$20M; StepFun's Step 5 Preview weights (around October 15), MiniMax-M3.1-Flash (October 23), and DeepSeek V4.1-Flash-1017 (October 17, MIT) also landed. Benchmarks reset again: Artificial Analysis Intelligence Index v5.0 (October 27) is not comparable with v4.x, puts Gemini 4 Carbon (59, not publicly available) and Claude Opus 5.5 (58) on top, and has the best open-weight models at about 32. On the tbench.ai Terminal-Bench 4.0 leaderboard (entries added October 6), Opus 5.5 (64.85%) and Sonnet 5.5 (61.82%) now sit above GPT-6 Astra and GPT-6.1 Sol (tied at 58.18%). OpenAI's October 23 legacy shutdown happened on schedule, and Claude Sonnet 4.5 still retires November 30 on the Claude API (Bedrock keeps it to April 8, 2027).

Update (November 2 to 6): On November 2, Google deprecated gemini-2.5-pro and gemini-2.5-flash on the Gemini API, reversing September's "not deprecated"; from December 1 (scheduled) they route to gemini-3.8-flash and gemini-3.5-flash-lite and bill at those rates. Anthropic released message threads in beta the same day and on November 3 extended its text watermark to Claude Mythos 5, Opus 4.7, and Sonnet 4.6. On November 5, OpenAI launched GPT-6-Cyber ($20/$100, Daybreak Red plus a separate model approval and IP allowlisting), a day after shipping gpt-image-2.6-sunburst and -flare at GPT Image 2.5 prices. Moonshot published the Kimi K3.1 Pro weights on November 4 under a custom license with a MaaS clause, Tencent released the full Hy4 the same day with weights promised but no date or license, and Ricursive's Apache 2.0 Ripple-65B Preview landed with no independent score yet. Artificial Analysis shipped Intelligence Index v5.0.1 on November 3: Gemini 4 Carbon 60, Claude Opus 5.5 59, GPT-6 Astra 57, GPT-6.1 Sol 55, and DeepSeek V4.1 Flash 1017 (33) now tops AA's open-weights ranking alone only because AA still labels K3.1 Pro (46) proprietary despite its released weights. OpenAI's pricing page now dates GPT-5.6 Sol's promotional price to November 30, and Alibaba has not shipped Qwen4.


Model Categories

By Capability Level (November 2026)

TierCharacteristicsExamplesUse Case
FrontierHighest capability ceiling; $4/$20 to $10/$50, or gatedClaude Fable 5.1, GPT-6 Astra, Claude Opus 5.5 ($4/$20), Gemini 4 Argon ($4/$20 from December 1), Meta Muse Wonder ($4/$20); gated: Gemini 4 Carbon, GPT-6 Rosalind, GPT-6-Cyber, Claude Mythos 5.1; subscription only: Muse Wonder FullHardest reasoning, long-horizon agents, ceiling-bound coding
Workhorse ($1 to $2 input)Near-frontier agentic quality at mid-tier priceClaude Sonnet 5.5, GPT-6.1 Sol, GPT-6 Sol ($2/$10), GPT-6.1 Terra ($1/$5), Grok 4.7 ($2/$6), Muse Spark 1.3 ($1.25/$4.25)Production agents, coding at scale, the default routing tier
Fast/EfficientLowest cost per token, high volumeGPT-6 Luna, Claude Haiku 5.5 (under 100K-token prompts), Gemini 3.8 Flash, DeepSeek V4.1-FlashClassification, extraction, routing, high-volume streaming
Previous Generation (still served)Well characterized, often superseded within weeksClaude Opus 5 / Sonnet 5 / Fable 5, Claude Haiku 4.5, GPT-5.6 Sol / Terra / Luna, Claude Opus 4.8, Gemini 3.6 FlashWorkloads that value stable, already-evaluated behavior
Small/EdgePrivate, edge, single GPUGemma 4 (12B, E4B), Qwen3.8-27B, IBM Granite 4.2, Muse Glimmer 30BLocal privacy, on-device, single-GPU serving
Reasoning-HeavyExtended internal reasoning at high effortGPT-6 Astra (xhigh/max), Claude Fable 5.1, Claude Opus 5.5 at high effort, DeepSeek V4.1-Flash (effort 1-100)Math, code debugging, multi-step logic

The fast tier now has two prices to watch on the Anthropic side: Claude Haiku 5.5 matches GPT-6 Luna at $0.10/$0.50 below a 100,000-token prompt and costs 5x on the whole request above it, so prompt length decides which tier it really belongs to.

By Reasoning Mode (November 2026)

ModeCapabilityModelsUse Case
Minimal reasoningThe fastest path the model allowsGPT-6.1 Terra at none, GPT-6.1 Sol at low, GPT-6 Luna at its lowest effort, Gemini 3.8 Flash at low, Claude Haiku 5.5 at low, Claude Sonnet 5.5 at between_toolsChat, simple extraction
Always-on reasoningThinking cannot be switched offClaude Opus 5.5 and Fable 5.1 (adaptive thinking always on), GPT-6 Astra (no none effort; low to max)Math, code debugging, planning
Dial-controlledEffort or thinking level per requestGPT-6.1 Terra (none to max, default medium), GPT-6.1 Sol (low to max, default medium), Claude Opus 5.5 and Haiku 5.5 (default medium), Gemini 3.8 Flash (thinking_level), DeepSeek V4.1-Flash (1-100)Variable-complexity tasks

The off switch is going away on Anthropic's side. Sonnet 5.5 returns a 400 on thinking: {type: "disabled"}, Opus 5.5 cannot disable thinking, and Haiku 5.5 rejects budget_tokens. OpenAI is split: Astra has no none effort, but GPT-6.1 Terra does. Sampling knobs are gone almost everywhere: under Claude API version 2026-10-21 (October 21), temperature, top_p, and top_k return a 400 at any value, even the default, on every model (earlier versions reject non-default values on Claude 4.7 and later), and the eight new SDK and CLI majors (Python 2.0.0, TypeScript 1.0.0, Go 2.0.0, Java 3.0.0, C# 13.0.0, PHP 1.0.0, Ruby 2.0.0, and the ant CLI 2.0.0, all now Apache 2.0 instead of MIT) send that version by default, so a routine SDK upgrade in any language can break sampling code on older models too. Astra accepts no temperature, top_p, or logprobs, and Google deprecated those parameters in the Gemini API on July 21. Effort is now the control surface, and defaults move between versions (Opus 5 defaulted to high, Opus 5.5 to medium), so set it explicitly on every route.


Frontier Models (November 2026)

Claude Opus 5.5 (Anthropic) - September 2026

AttributeValue
Model IDclaude-opus-5-5 (Bedrock: anthropic.claude-opus-5-5)
Context Window1M tokens; 128K max output (300K on Batch with the beta header)
Input / Output Cost$4.00 / $20.00 per 1M (Opus 5: $5 / $25)
Cache / BatchCache write $5.00 (5 min) or $8.00 (1 hr); cache read $0.20 (0.05x input); Batch $2 / $10; minimum cacheable prompt 512 tokens
Fast mode$8 / $40 per 1M, out of research preview since October 15; on the Claude API, Claude Models on Google Cloud (Anthropic managed), and Foundry's Azure-hosted deployments; not on Batch. Down from $10 / $50 on Opus 5 and Opus 4.8
ThinkingAdaptive thinking always on (cannot be disabled); default effort medium (Opus 5 defaulted to high)
AvailabilityClaude API, Amazon Bedrock, Google Cloud (Google managed or Anthropic managed), Microsoft Foundry, Claude Platform on AWS; zero data retention available
BenchmarksAnthropic-reported: Terminal-Bench 4.0 66.4% at xhigh (SE ±2.6) vs Fable 5.1 55.8% (Anthropic's run) and GPT-6 Astra 57.9% (OpenAI's figure, high effort); OSWorld 2.1 partial 81.8%; GDPval-AA v2.1 1846 Elo; HLE with tools 67.7%. Independent: Terminal-Bench 4.0 leaderboard 64.85% at max, first (tbench.ai, entries added October 6); Artificial Analysis Intelligence Index v5.0.1 59 (max; 58 on v5.0, with fallback), second only to Gemini 4 Carbon (60), which developers cannot call
ReleasedSeptember 22, 2026 (knowledge cutoff June 2026; retirement not sooner than September 22, 2027)

What it is: The second Opus price cut (after Opus 4.5 came down from $15/$75), and the model Anthropic's docs now tell you to start with for most workloads. Fable 5.1 is reserved for demanding reasoning, long-horizon agentic work, or cases where Opus 5.5 at higher effort still falls short in your evals. A cheaper model that beats the more expensive Fable 5.1 on most of Anthropic's own agentic benchmarks inverts the usual "pay more for the ceiling" routing logic. Anthropic claims about 40% lower cost than Opus 5 on typical workloads (vendor-reported).

Migration traps from Opus 5: the default effort dropped from high to medium, so an unpinned upgrade changes both latency and quality; forced tool_choice (any or tool) returns a 400; thinking blocks are bound to the producing model and conversation; and on the Claude API and Google Cloud only computer_toolset_20260801 is accepted for computer use. Opus 5.5 adds a biology safety classifier alongside the cyber one. When safeguards intervened during Anthropic's evals, cyber tasks were completed by Opus 4.8 and biology and frontier-LLM-development tasks by Opus 5, so the headline scores belong to a routed system, not one set of weights.

Claude Sonnet 5.5 (Anthropic) - September 2026

AttributeValue
Model IDclaude-sonnet-5-5
Context Window1M tokens; 128K max output
Input / Output Cost$2.00 / $10.00 per 1M (unchanged from Sonnet 5)
Cache / BatchCache write $2.50 (5 min) or $4.00 (1 hr); cache read $0.10 (0.05x, cut from $0.20 on October 7); Batch $1 / $5
ThinkingAdaptive thinking on by default; the lowest setting is between_tools (accepted at high effort or below), and thinking: {type: "disabled"} returns a 400. Default effort high on the API, medium in Claude Code and the Claude apps
AvailabilityClaude API, Bedrock, Google Cloud, Foundry, Claude Platform on AWS; zero data retention available
BenchmarksAnthropic-reported, with Opus 5.5 in parentheses: Terminal-Bench 4.0 70.6% (66.4%), CursorBench 4.0 55.5% (57.8%), GDPval-AA v2.1 1844 (1846), OSWorld 2.1 partial 80.1% (81.8%), HLE with tools 64.5% (67.7%), FrontierCode 1.1 46.2% at max effort (54.4%); Anthropic notes Sonnet 5.5 scored lower at max than at xhigh on this test. Independent: Terminal-Bench 4.0 leaderboard 61.82% at max, second behind Opus 5.5 (tbench.ai, entries added October 6); AA Intelligence Index v5.0 54 (max, with fallback), 50 at xhigh
ReleasedSeptember 28, 2026 (knowledge cutoff June 2026; retirement not sooner than September 28, 2027)

What it is: A $2/$10 model within a few points of the flagship on most of Anthropic's own vendor-reported agentic evals (about 8 points behind on FrontierCode 1.1), which collapses the case for sending most agent traffic to Opus. The independent picture is less flat: AA v5.0 has it 4 points behind Opus 5.5 at max effort, and the gap widens at lower effort (41 vs 52 at high), so the effort you serve matters more than the list price. The runner flips the order too: Anthropic's own Terminal-Bench 4.0 runs put Sonnet 5.5 ahead of Opus 5.5 (70.6% vs 66.4%), while the tbench.ai leaderboard puts it behind (61.82% vs 64.85%). Anthropic claims 30%+ faster output than Sonnet 5 and up to 30% lower cost per task (vendor-reported), and the October 7 cache-read cut to 0.05x makes long agent loops cheaper still. Anthropic's Sonnet 5.5 table puts Sonnet 5 at 10.3% on Terminal-Bench 4.0, while its Opus 5.5 table puts Opus 5 at 52.3%; treat the Sonnet 5 figure as an outlier and do not use it as a baseline.

Breaking changes vs Sonnet 5: disabling thinking returns a 400 (send between_tools instead); forced tool use returns a 400; thinking blocks are bound to the model, the conversation, and the producing account (blocks sent from an unlinked account are dropped and recorded as organization_binding_mismatch); computer_20251124 is rejected on the Claude API and Google Cloud; and temperature, top_p, and top_k return a 400 (at non-default values on older API versions, at any value on 2026-10-21). Sonnet 5 became a legacy model three months after launch.

Claude Fable 5.1 and Mythos 5.1 (Anthropic) - September 2026

AttributeValue
Model IDsclaude-fable-5-1 (GA; Bedrock anthropic.claude-fable-5-1); claude-mythos-5-1 (limited availability through the Cyber Verification Program, which absorbed Project Glasswing on October 6)
RelationshipSame weights, different safeguards; Mythos 5.1 is the more permissive variant
Input / Output Cost$10.00 / $50.00 per 1M (unchanged from Fable 5)
Cache / BatchCache write $12.50 (5 min) or $20.00 (1 hr); cache read $0.25 (0.025x input, down from $1.00 on Fable 5); Batch $5 / $25
Context Window1M tokens at standard pricing; 128K max output
ThinkingAdaptive thinking always on; default effort high
Data handling30-day retention required; not available under zero data retention unless Anthropic expressly authorizes it (customers eligible for Enterprise Frontier Safeguards can use Fable 5.1 and Fable 5 with ZDR until EFS is ready)
ProvenanceAnthropic's statistical text watermark on outputs on every platform; signed C2PA Content Credentials on generated media retrieved via the Files API. Since November 3, text from Claude Mythos 5, Opus 4.7, and Sonnet 4.6 carries the watermark too, but on Sonnet 4.6 text generated with a non-default temperature, top_p, or top_k might not, so check model and sampling settings before relying on vendor marking for disclosure duties
BenchmarksTerminal-Bench 4.0: 57.88% at max on the tbench.ai leaderboard, 55.8% in Anthropic's own run. AA Intelligence Index v5.0: 50 (xhigh, with fallback), 49 at max, below Opus 5.5's 58
ReleasedSeptember 1, 2026 (knowledge cutoff June 2026; retirement not sooner than September 1, 2027)

What changed: the cost of staying on the ceiling model for long agent loops. Anthropic says the 75% cache-read cut lowers typical bills about 25% and highly agentic ones up to 45%, and that its newest cyber safeguards produce about 60% fewer false positives than Fable 5's earlier ones (both vendor-reported). Fable 5.1 may now be used to discover vulnerabilities but not to develop exploits. Fable 5 remains available.

Mythos 5.1 access: on October 6 Anthropic expanded the Cyber Verification Program into three tiers (Defense, Red Team, Specialized), folded Project Glasswing into it, and made Mythos 5.1 available in every tier. Data retention is required; customers who already hold zero-retention access to Fable 5.1 or Mythos 5.1 keep it until Enterprise Frontier Safeguards ships. Anthropic also said in September that Mythos 5.1 would be offered through the Life Sciences Verification Program (first participants enrolled with the US government). The design consequence: the permissive variant is now a verification-tier product with its own retention terms, not a partner list.

Safeguard fallback is now an API contract. On Fable 5 a classifier hit handed the request to Opus 4.8 automatically. On Fable 5.1 it is explicit and configurable:

Rendering diagram...
Click to expand and download

From September 24, refusals that arrive before any output in the bio, frontier_llm, and reasoning_extraction categories are billed on all platforms (other categories stay unbilled); mid-stream refusals bill the input plus the output already streamed; and fallback credit refunds the prompt-cache cost of switching models. Budget for refusals, and route on stop_reason, not HTTP status.

Claude 5.5 / 5.1 Migration Checklist

ChangeApplies toWhat breaksDo this instead
Forced tool_choice (any or tool) returns 400Fable 5.1, Mythos 5.1, Opus 5.5, Sonnet 5.5 (Haiku 5.5 accepts it, but the response starts with the tool call and has no thinking block)Structured extraction that forces a tool callauto plus strict tool use, or structured outputs
Thinking blocks bound to model and conversationFable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 (Mythos 5.1 skips the prefix check); Sonnet 5.5 and Haiku 5.5 also bind blocks to the producing accountReplaying history to another model silently drops reasoning (Haiku 5.5 drops blocks from Opus 5, Opus 5.5, Sonnet 5.5, Fable, and Mythos without an error); editing anything before a thinking block returns 400 for accounts created on or after August 31, 2026Treat history as append-only and replay it through the producing account; opt into prefix_mismatch_behavior: "drop_block" with the thinking-binding-controls-2026-08-01 beta header
Thinking cannot be disabledOpus 5.5; Sonnet 5.5's floor is between_toolsLatency budgets that assumed thinking offSet effort explicitly and re-measure latency
Default effort changedOpus 5.5 is medium (Opus 5 was high)Silent quality and latency shift on upgradePin effort on every request
computer_20251124 rejectedOpus 5.5 and Sonnet 5.5 on the Claude API and Google Cloud (Bedrock still accepts it)Computer-use agents on the old tool versionMove to computer_toolset_20260801
Pre-output refusals billed in three categoriesAll platforms since September 24Cost models that treat refusals as freeTrack stop_reason: "refusal" as a cost line
Sampling parameters return 400 at any value; document blocks in assistant messages and citations on user text blocks return 400Every model under API version 2026-10-21, which all eight new SDK and CLI majors (Python 2.0.0, TypeScript 1.0.0, Go 2.0.0, Java 3.0.0, C# 13.0.0, PHP 1.0.0, Ruby 2.0.0, ant CLI 2.0.0) send by default; on Bedrock the version works only for Claude 4.7 and laterCode that set temperature or top_p explicitly, even to defaults, on any Claude modelDrop the parameters, then pin the SDK major and API version so the upgrade is a deliberate change
Haiku 4.5 to Haiku 5.5Fast-tier migrationsbudget_tokens, sampling parameters, and assistant prefill return 400; about 30% more tokens for the same text, and earlier turns' thinking blocks now stay in context as input; the high-resolution image tier bills a 2000x1500 image at about 2.5x the visual tokens; prompts over 100K tokens bill 5x on the whole request; thinking blocks are account-bound and prefix-checked (Haiku 4.5 ran neither check); no Priority TierReplace prefill with structured outputs, re-count text and image tokens with the new model, route long prompts elsewhere, keep history append-only, and plan capacity separately from any Haiku 4.5 Priority Tier commitment

Two smaller changes trip harnesses: text between tool calls now arrives in thinking blocks that are empty at the default display: "omitted" (the display: "updates" beta returns the text), and supporting betas shipped for per-message effort and turn-scoped system messages (September 1), compaction on demand (compact-2026-09-04, September 14), and inline tool definitions mid-conversation without cache loss (inline-tools-2026-09-15, September 22). On November 2 Anthropic added message threads (beta, message-threads-2026-08-12): the server stores the conversation, system prompt, and tools, and later requests send only new messages. Threads are best-effort, expire 5 minutes after the last request, can return 404 thread_not_found, are not ZDR eligible, and do not run on Bedrock or Google-managed Vertex, so keep your own copy of history and replay it; the thread is a cache, not the system of record.

Claude Haiku 5.5 (Anthropic) - October 2026 NEW

AttributeValue
Model IDclaude-haiku-5-5 (Bedrock: anthropic.claude-haiku-5-5)
Context Window1M tokens; 128K max output
Price, prompt up to 100,000 tokens$0.10 / $0.50 per 1M; cache write $0.125 (5 min) or $0.20 (1 hr); cache read $0.01; Batch $0.05 / $0.25
Price, prompt over 100,000 tokens$0.50 / $2.50 per 1M; cache write $0.625 or $1.00; cache read $0.05; Batch $0.25 / $1.25. Applies to the whole request, and prompt length counts cache reads and cache writes
ThinkingAdaptive thinking, effort low to max, default medium; budget_tokens returns a 400. Unlike Opus 5.5 and Sonnet 5.5, it accepts a forced tool_choice, but the response then starts with the tool call and has no thinking block
Breaking changes vs Haiku 4.5budget_tokens, temperature, top_p, top_k, and assistant prefill return 400; the newer tokenizer produces about 30% more tokens for the same text; the high-resolution image tier costs about 2.5x the visual tokens for a 2000x1500 image; thinking blocks are bound to the producing account and prefix-checked; Priority Tier is not supported
AvailabilityClaude API, Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS
BenchmarksAnthropic-reported: Terminal-Bench 4.0 39.2%; OSWorld 2.1 offline subset 72.4% (partial credit). Independent: AA Intelligence Index v5.0 36 (max)
SafetySystem card: Gray Swan prompt-injection attack success within 15 attempts 7.1% overall and 24.4% on GUI computer use
ReleasedOctober 7, 2026 (retirement not sooner than October 7, 2027)

What it is: Anthropic's fast tier at GPT-6 Luna's list price, with an independent score well above Luna's (AA v5.0: 36 vs 28, both at max effort). It is also the first Claude 4.6-or-later model with a long-context price cliff, and the cliff is harsher than OpenAI's. On GPT-6 Luna, crossing 272K input doubles input and multiplies output by 1.5; on Haiku 5.5, crossing 100K multiplies every token type by 5. A 99K-token prompt with 1K of output costs about $0.0104; a 101K-token prompt with the same output costs about $0.053. Because cache reads count toward the threshold, a cached 120K-token prefix still bills every request at the higher rate.

Three design consequences. First, prompt-length routing is now a cost control on Anthropic, not just on OpenAI and xAI: keep Haiku 5.5 traffic under 100K, and send longer prompts to Sonnet 5.5 (flat $2/$10 to 1M) or GPT-6 Luna (long-context $0.20/$0.75). Second, the tokenizer moves the line: a prompt that measured 80K tokens on Haiku 4.5 is roughly 104K on Haiku 5.5, and thinking blocks from earlier turns now stay in context as input (Haiku 4.5 kept only the latest turn's), so multi-turn conversations reach the cliff sooner. Re-count with the new model before trusting old prompt-size dashboards. Third, the prompt-injection numbers are high for a model you might put on a browser or desktop: about one GUI computer-use attack in four succeeds within 15 attempts, so keep Haiku 5.5 agents behind the same isolation and human-approval gates as larger models.

Claude Haiku 4.5 (Anthropic) - October 2025, Superseded by Haiku 5.5

Still Active on the Claude API with no deprecation notice as of November 6. Its floor ("not sooner than October 15, 2026") has passed, and with Anthropic's 60-day minimum notice it cannot retire there before about January 2027. The consumer apps moved first: on October 29 Anthropic removed Haiku 4.5 from the model selector in Claude chat, Cowork, and Design for Free, Pro, and Max users and moved those chats to Haiku 5.5, while Claude Code, Team and Enterprise plans, and the API keep it. Microsoft's Foundry schedule still lists claude-haiku-4-5 for November 15, 2026, nine days out, although Anthropic says its dates apply to Foundry, so plan for the earlier date on Foundry now and confirm with both vendors. At $1/$5 with a 200K window it costs 10x Haiku 5.5 below 100K tokens; the remaining reasons to stay are behavioral (assistant prefill, budget_tokens, the older tokenizer) or contractual (a Priority Tier commitment, which Haiku 5.5 does not support), so treat them as migration work rather than a reason to keep paying 10x.

Claude Opus 5 (Anthropic) - July 2026, SUPERSEDED by Opus 5.5

AttributeValue
Model IDclaude-opus-5
Context Window1M tokens (default and max; 128K max output)
Input / Output Cost$5.00 / $25.00 per 1M (unchanged from Opus 4.8)
Fast mode$10.00 / $50.00 per 1M, about 2.5x faster
BenchmarksAnthropic's launch post reported the pre-release benchmark it called Frontier-Bench v0.1 (later Terminal-Bench 3.0, itself replaced by Terminal-Bench 4.0 on August 28; scores do not carry over): 43.3% at max effort vs GPT-5.6 Sol's 34.4%, Fable 5's 33.7%, and Opus 4.8's 18.7%; within 0.5% of Fable 5 on CursorBench 3.2 at about half the cost per task. Vendor-reported. Since then: Terminal-Bench 4.0 leaderboard 53.94% at xhigh; SWE-Bench Pro v2 (Scale, September 22) 99.4% on the saturated public split and 81.6% (222/272) on the private set, the top private score Scale published
ReleasedJuly 24, 2026 (Claude API, Claude Code, Claude Cowork; new default on Claude Max)

What it is: The Opus line's generational successor at unchanged pricing, aimed at long-horizon agentic coding and computer use. Beta features include mid-conversation tool changes and automatic fallback routing. The dual-price Fast mode continues the pattern Opus 4.8 introduced: one model, two latency tiers.

Status: Superseded by Opus 5.5 on September 22 at a lower price ($4/$20 vs $5/$25). Still available as a legacy model, and one of Fable 5.1's two permitted safeguard fallback targets (with Opus 4.8).

Claude Sonnet 5 (Anthropic) - June 2026, SUPERSEDED by Sonnet 5.5

AttributeValue
Model IDclaude-sonnet-5
Context Window1M tokens per third-party coverage (not stated in the launch post)
Input / Output Cost$2.00 / $10.00 per 1M (permanent since August 10, 2026)
Cache / BatchCache write $2.50 per 1M (5 min) or $4.00 (1 hr); cache hit $0.20; Batch API $1.00 / $5.00
PositioningThe most agentic Sonnet at launch: planning, browser and terminal tool use, autonomous operation approaching Opus 4.8 at lower cost
Safety postureCyber safeguards on by default; deliberately reduced cybersecurity capability relative to Opus-class models
ReleasedJune 30, 2026 (default model across consumer and developer products same day)

What it is: The production workhorse that replaced Sonnet 4.6 as the default. On August 10, 2026 Anthropic made the introductory $2/$10 rate permanent and canceled the September 1 increase to $3/$15, so Sonnet 5 is permanently cheaper than the Sonnet 4.6 it succeeded (still $3/$15).

Status: Superseded by Sonnet 5.5 on September 28 at the same price; still available as legacy. Separately, Anthropic deprecated Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) on September 30, with retirement on November 30, 2026 on the Claude API and Foundry and claude-sonnet-5-5 as the named replacement. Bedrock runs its own clock: it moved Sonnet 4.5 to Legacy on October 8, with public extended access from January 8, 2027 and end of life April 8, 2027.

Claude Fable 5 (Anthropic) - June 2026, SUPERSEDED by Fable 5.1

AttributeValue
Model IDclaude-fable-5
Context Window1M tokens (Opus 4.7 tokenizer; roughly 30% more tokens than pre-4.7 models for the same text)
Max Output128K tokens
Input Cost$10.00 / 1M tokens
Output Cost$50.00 / 1M tokens
ThinkingAdaptive thinking, always on (no separate extended-thinking toggle)
MultimodalText + Vision (new state of the art on vision tasks per Anthropic)
BenchmarksState of the art on nearly all tested benchmarks per Anthropic; highest frontier score on Cognition's FrontierCode, highest on the Hebbia Finance Benchmark, ViBench, and CursorBench. Standard numeric scores (SWE-bench, GPQA) were not published in the launch post.
ReleasedJune 9, 2026 (GA on Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry)

What it is: A Mythos-class model made safe for general availability. Until then the Mythos line (SWE-bench Verified 93.9% on Mythos Preview) was restricted to ~11 Project Glasswing partners over dual-use cybersecurity concerns. Fable 5 brought that capability tier to everyone by pairing it with conservative safeguards.

The Opus 4.8 fallback safeguard: When Fable 5's classifiers detected a request in one of three categories (offensive cyber techniques, bioweapon-adjacent biology and chemistry, or attempts to distill the model), the response was delegated to Claude Opus 4.8 and the user was informed. Anthropic said this triggered in under 5% of sessions and was deliberately tuned conservative, so some harmless requests got caught. Architecturally this is a production example of model-tier routing as a safety control, not just a cost control. Fable 5.1 turned it into the explicit, billable API contract described above.

Considerations: 2x the per-token price of Opus 4.8 ($10/$50 vs $5/$25). Mythos-class traffic carries a 30-day data retention requirement (not used for training; access-logged; deleted after 30 days in almost all cases), which matters for compliance reviews. On subscription plans it was included at no extra cost June 9-22, then moved to usage credits. Fable 5 cache reads were $1.00 per 1M; Fable 5.1 cut them to $0.25 at the same list price, so there is little reason to start new work on Fable 5.

Claude Mythos 5 (Anthropic) - June 2026, SUPERSEDED by Mythos 5.1

AttributeValue
Model IDclaude-mythos-5
StatusLimited availability at launch: Project Glasswing partners and select biology researchers
RelationshipSame underlying model as Fable 5 with safeguards lifted in some areas
Pricing$10 / $50 per 1M (less than half of Mythos Preview)
ReleasedJune 9, 2026

Why it matters: Succeeded Claude Mythos Preview at comparable or somewhat stronger capability and much lower price. The Fable/Mythos split formalized a two-track release pattern: one safeguarded general release, one unrestricted release for vetted defenders. Mythos 5.1 continues it.

Claude Opus 4.8 (Anthropic) - May 2026, Legacy

AttributeValue
Context Window1M tokens (standard pricing across the full window)
Input Cost$5.00 / 1M tokens (same as 4.7)
Output Cost$25.00 / 1M tokens (same as 4.7)
Cache: 5m write$6.25 / 1M tokens
Cache: 1h write$10.00 / 1M tokens
Cache: hit / refresh$0.50 / 1M tokens
Batch API$2.50 / $12.50 per 1M (50% discount)
Fast mode$10 / $50 per 1M in research preview (about 2.5x faster; 3x cheaper than the Opus 4.7 fast mode, which was $30 / $150). Deprecated October 30; removed November 30, 2026, after which speed: "fast" returns an error
Extended ThinkingNative, adaptive mode
MultimodalText + Higher-resolution Vision
SWE-bench Verified88.6%
SWE-Bench Pro69.2% (up from 64.3% on Opus 4.7)
Terminal-Bench 2.174.6% (GPT-5.5 led at 78.2% at release)
GDPval-AA1890 Elo (up from 1753 on Opus 4.7)
OSWorld-Verified82.3%
Online-Mind2Web84%
ReleasedMay 28, 2026 (GA on Claude API, AWS Bedrock, Vertex AI)

Best for: Long-running autonomous coding work in Claude Code, codebase-scale migrations, agentic workflows that need parallel subagents, and workloads where the alignment and honesty gains matter. It is also one of Fable 5.1's two permitted safeguard fallback targets (with Opus 5), and Anthropic's Opus 5.5 evals routed cyber tasks to it when safeguards intervened, so it stays in production paths even for teams that never call it directly.

Key features over Opus 4.7:

  • Dynamic Workflows (research preview): Claude plans the work and runs hundreds of parallel subagents in a single Claude Code session, verifies their outputs, and reports back. Suited for codebase-scale migrations across hundreds of thousands of lines.
  • Mid-task system messages: The Messages API now accepts system messages mid-conversation, useful for steering long agent runs without ending the session.
  • Optional fast mode at roughly 2.5x speed for $10 / $50 per 1M, priced 3x lower than the Opus 4.7 fast mode (removed November 30, 2026; Opus 5.5 fast mode at $8 / $40 is the replacement).
  • Effort-control toggle in claude.ai and Cowork lets users tune reasoning depth per turn.
  • Expanded Claude Code rate limits.

Considerations: Tokenizer is the same one introduced in Opus 4.7: roughly 30% more tokens than the pre-4.7 tokenizer for the same text, per Anthropic's token-counting docs, with the exact increase depending on content (the Opus 4.7 launch post gave a range of 1.0x to 1.35x). At release, GPT-5.5 held the SWE-Bench Verified leaderboard at 88.7% and led Terminal-Bench 2.1 at 78.2%. GPQA Diamond slipped 0.6 pts versus Opus 4.7. Anthropic's tokenizer change means token counts and bills for the same text are not directly comparable to pre-4.7 models. There was no Claude Sonnet 4.8; the line jumped to Claude Sonnet 5 on June 30, 2026, which replaced Sonnet 4.6 as the production workhorse. For new work, Opus 5.5 is cheaper ($4/$20) and stronger.

[!NOTE] Retired August 5, 2026: claude-opus-4-1-20250805 was removed from the Claude API, closing out the last $15/$75 per 1M Opus tier. At that point every first-party Opus SKU was $5/$25 ($10/$50 in Fast mode); Opus 5.5 later cut the current Opus to $4/$20. Amazon Bedrock and Google Cloud set their own schedules: on Bedrock, Opus 4.1 moved into higher-priced public extended access on October 8, 2026 and reaches end of life on January 8, 2027. Code pinned to that ID fails on the first-party API while still working on the partner clouds.

Claude Opus 4.7 (Anthropic) - April 2026, Legacy

AttributeValue
Context Window1M tokens
Max Output128K tokens
Input Cost$5.00 / 1M tokens (same as 4.6)
Output Cost$25.00 / 1M tokens (same as 4.6)
Extended ThinkingNative, Adaptive mode
MultimodalText + Higher-resolution Vision
SWE-bench Verified (Adaptive)87.6% (May 13, 2026)
ReleasedApril 16, 2026 (GA on API, Bedrock, Vertex, Microsoft Foundry)

Status: Still available as a legacy model. It powered Claude Code at release; Claude Code's default Opus has been Opus 5.5 since 2.1.280. Prefer Opus 5.5 for new work: it is cheaper and stronger.

Claude Mythos Preview (Anthropic) - SUCCEEDED BY MYTHOS 5 AND 5.1

AttributeValue
StatusRestricted research preview, Project Glasswing partners only (~11 orgs: AWS, Apple, Cisco, Google, Microsoft, NVIDIA, Palo Alto, etc.); deprecated since June 9, 2026, retirement date to be announced
Reason for restrictionDual-use cybersecurity capabilities
SWE-bench Verified93.9% (May 13, 2026; the published SOTA before the Fable 5 / Mythos 5 launch)
ReleasedApril 7, 2026 (restricted partner preview); succeeded by Claude Mythos 5 on June 9, 2026 at less than half the price

Best for: Historical reference. Its capability tier reached general availability as Claude Fable 5 on June 9, 2026; new gated security work should target Mythos 5.1 through the Cyber Verification Program, which absorbed Glasswing on October 6.

Claude Opus 4.6 (Anthropic) - February 2026, Legacy

AttributeValue
Context Window1M tokens
Max Output128K tokens
Input Cost$5.00 / 1M tokens
Output Cost$25.00 / 1M tokens
Extended ThinkingNative adaptive thinking (configurable budget_tokens)
MultimodalText + Vision
HighlightsAnthropic's flagship at release; now legacy
ReleasedFebruary 2026

Status: Still available. Opus 5.5 is cheaper ($4/$20 vs $5/$25) and stronger, so there is no cost reason to stay on it.

Claude Sonnet 4.6 (Anthropic) - February 2026, Legacy

AttributeValue
Context Window1M tokens
Input Cost$3.00 / 1M tokens
Output Cost$15.00 / 1M tokens
Extended ThinkingSupported
MultimodalText + Vision
HighlightsHandled tasks that previously required the Opus tier; the cost/quality default through mid-2026
ReleasedFebruary 2026

Status: Still available, but at $3/$15 it now costs more than Sonnet 5.5 ($2/$10). Migrate unless your evals say otherwise.

GPT-6 Astra (OpenAI) - September 2026

AttributeValue
Model IDgpt-6-astra
Context Window1.05M tokens; 128K max output; knowledge cutoff April 30, 2026
Input / Output Cost$10.00 / $50.00 per 1M; above 272K input the whole request bills at $20 / $75
CacheCached input $1.00; cache write $12.50 (30-minute TTL) or $25.00 (12-hour TTL, Responses API, since October 20)
Service tiersBatch and Flex 50% off; Fast 2x ($20 / $100); Ultrafast 6x ($60 / $300, GA September 29, Responses API)
ReasoningEfforts low, medium, high, xhigh, max; no none; no custom temperature, top_p, or logprobs; tool calling requires the Responses API
AvailabilityStaged rollout from September 3 (Daybreak organizations first), in the API by September 5; Amazon Bedrock GA September 8 (In-Region and Geo inference $11 / $55, 10% over Global); Microsoft Foundry GA; Google Cloud Model Garden private preview for select enterprise customers since October 22 (no GA date or per-token prices published)
Safety ratingFirst OpenAI model rated Critical for cybersecurity under its Preparedness Framework; High for biological and chemical capability. OpenAI says covert sandbagging is likely not reliably caught
BenchmarksTerminal-Bench 4.0 leaderboard 58.18% at max, tied with GPT-6.1 Sol behind Claude Opus 5.5 (64.85%) and Sonnet 5.5 (61.82%) (tbench.ai, entries added October 6). ARC-AGI-3 (ARC Prize): 62.7% on the provider-neutral Standard harness (max, $26,098) vs 99.9% on ARC Prize's Provider Adapter harness (high, $18,817), which uses OpenAI's own context-management features. AA Intelligence Index v5.0 and v5.0.1: 57 (max)
ReleasedSeptember 3, 2026

What it is: OpenAI's new ceiling model, at the same $10/$50 as Claude Fable 5.1. Three things change cost models. First, speed is a priced axis with three rungs (Standard, Fast at 2x, Ultrafast at 6x), so a latency SLO is a budget decision. Second, the 272K cliff applies to the entire request, so a prompt just over the line costs roughly twice one just under it: prompt-length routing is a cost control. Third, token efficiency: Artificial Analysis measured Astra using about a third of GPT-5.6 Sol's tokens per coding task, so its cost per task sits closer to the mid tier than the list price suggests.

Read the harness: the 37-point ARC-AGI-3 swing on the same model and test set comes from whether the scaffold preserves opaque reasoning state between requests. Quote Astra's numbers with harness, effort, and who ran them.

GPT-6 Sol and GPT-6 Luna (OpenAI) - September 2026

AttributeValue
Model IDsgpt-6-sol, gpt-6-luna (text and image in, text out; Responses and Chat Completions)
Context Window1.05M tokens; 128K max output
Sol pricing$2.00 / $10.00 per 1M; cached input $0.20; cache write $2.50; above 272K $4 / $15
Luna pricing$0.10 / $0.50 per 1M; cached input $0.01; cache write $0.125; above 272K $0.20 / $0.75
Batch / FlexSol $1 / $5; Luna $0.05 / $0.25
LineageSame types of data and training as GPT-6 Astra, per OpenAI. They replace GPT-5.6 Sol and Luna at about half the price; the Terra tier followed as GPT-6.1 Terra on October 26
AvailabilityAPI and Amazon Bedrock GA September 22, 2026; GPT-6 Luna (not Sol) is also in Google Cloud Model Garden private preview since October 22
Known issueAn image-encoding bug degraded image understanding and computer use until a September 25 fix shipped under the same model IDs; OpenAI told customers to rerun image evals
ReleasedSeptember 22, 2026

Why it matters: Luna at $0.10/$0.50 is OpenAI's cheapest current-generation model, and since October 7 Claude Haiku 5.5 matches that list price below a 100K-token prompt (AA v5.0 scores Luna 28 and Haiku 5.5 36, both at max). The retiring gpt-5-nano ($0.05/$0.40) stays cheaper until it shuts down December 11, and OpenAI names gpt-5.6-luna ($0.20/$1.20) as its replacement, so moving to Luna instead avoids a price rise. Luna undercuts DeepSeek V4.1-Flash even off-peak ($0.15/$0.60). Gemini 2.5 Flash-Lite ($0.10/$0.40, existing users only) and Meta's Muse Spark contributor tier ($0.10/$0.20) are cheaper, and some hosted open models are cheaper on output (MiMo-V2.6-Flash $0.14/$0.28, Qwen3.8-Flash $0.15/$0.47). GPT-6.1 Terra (October 26, $1/$5) restored the middle rung, so three-tier Luna, Terra, Sol routing is back. The image bug is the operational lesson: a provider-side fix changed behavior under a stable ID, so evals must rerun on vendor changelog events, not only on model-ID changes. OpenAI's Sol post reports GPT-6 Sol at xhigh 60.5% vs Claude Opus 5 at medium 60.3% on its offline OSWorld 2.0 set (vendor-reported).

GPT-6.1 Terra (OpenAI) - October 2026 NEW

AttributeValue
Model IDgpt-6.1-terra (on Bedrock, call it through the us.openai.gpt-6.1-terra or global.openai.gpt-6.1-terra inference profile)
Context Window1.05M tokens
Input / Output Cost$1.00 / $5.00 per 1M; above 272K input the whole request bills at $2.00 / $7.50
CacheCached input $0.05 (0.05x); cache write $1.25 (30-minute TTL) or $2.50 (12-hour)
Service tiersBatch and Flex $0.50 / $2.50; Fast $2 / $10; no Ultrafast
Reasoningnone through max, default medium (Astra has no none)
AvailabilityAPI; Amazon Bedrock
Safety ratingSystem card: Critical for cybersecurity, High for biological capability
BenchmarksAA Intelligence Index v5.0: 51 (max), 38 at the default medium
ReleasedOctober 26, 2026

What it is: the middle rung OpenAI skipped in September, at half the price of GPT-6.1 Sol and Claude Sonnet 5.5 and half the GPT-5.6 Terra ($2/$12) it replaces. On AA v5.0 it sits 4 points behind GPT-6.1 Sol and 3 behind Sonnet 5.5, all at max effort. The catch is effort: AA scores it 38 at the default medium and 51 at max, so the headline number costs max-effort reasoning tokens. Like Astra, it is rated Critical for cybersecurity under OpenAI's Preparedness Framework, so check its usage terms before routing security work to it. For routers it is the obvious "try Terra before Sol" step, and a cheaper default than Sol for most agent traffic if it passes your evals.

GPT-6.1 Sol (OpenAI) - September 2026

AttributeValue
Model IDgpt-6.1-sol
Context Window1.05M tokens; 128K max output; knowledge cutoff April 30, 2026
Input / Output Cost$2.00 / $10.00 per 1M; above 272K $4 / $15
CacheCached input $0.10 (0.05x, half of GPT-6 Sol's rate); cache write $2.50 (30-minute) or $5.00 (12-hour)
Speed tiersFast $4 / $20; Ultrafast $12 / $60 ($24 / $90 above 272K), for all API users since October 8
Reasoninglow, medium (default), high, xhigh, max
FeaturesHosted shell, computer use, MCP, multi-agent delegation (beta, Responses API)
AvailabilityAPI; Amazon Bedrock, where it follows OpenAI's lifecycle and deprecation terms rather than a Bedrock minimum-availability date; Google Cloud Model Garden private preview since October 22
BenchmarksTerminal-Bench 4.0 leaderboard 58.18%, tied with GPT-6 Astra (tbench.ai, entries added October 6). AA Intelligence Index v5.0 and v5.0.1: 55 (max)
ReleasedSeptember 29, 2026

What it is: OpenAI's model for complex coding and professional work at a lower cost than Astra, shipped one week after GPT-6 Sol at the same list price. On AA v5.0.1 it trails Astra by 2 points and Opus 5.5 by 4, all at max effort, and it ties Astra on the Terminal-Bench 4.0 leaderboard, at a fifth of Astra's list price. Codex CLI switched its default to GPT-6.1 Sol in rust-v0.159.1 the same day, a silent default swap for anyone who had not pinned a model. Note the Bedrock terms: cross-cloud availability no longer implies a cloud-guaranteed support window. With Google Cloud's October 22 private preview of GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna (billed through Google Cloud Marketplace and drawing down existing commitments), all three hyperscalers now resell OpenAI frontier models, so committed-spend drawdown becomes a routing input; a preview carries no SLA, and Google lists Confidential Compute as "coming soon".

GPT-5.4 (OpenAI) - March 2026

AttributeValue
Context Window272K tokens (standard); extended available
Input Cost$2.50 / 1M tokens (cached $0.25)
Output Cost$15.00 / 1M tokens
MultimodalText, Vision, native computer use
HighlightsBuilt-in computer-use capabilities; 33% fewer factual errors vs GPT-5.2; combines coding + agentic strengths
ReleasedMarch 2026

Best for: Existing agentic workflows already validated on it. Considerations: Above 272K input tokens the whole request bills at long-context rates ($5 / $22.50). GPT-6 Sol is cheaper ($2/$10) for new builds.

GPT-5.4-mini (OpenAI) - March 2026

AttributeValue
Context Window272K tokens
Input Cost$0.75 / 1M tokens
Output Cost$4.50 / 1M tokens
HighlightsCost/performance option for high-volume GPT-5 tier workloads
ReleasedMarch 2026

Best for: High-volume API calls already tuned on it. For new volume tiers, GPT-6 Luna ($0.10/$0.50) is far cheaper. GPT-5.4-nano is $0.20 / $1.25 and retires April 1, 2027.

GPT-5.4 Pro (OpenAI) - March 2026

AttributeValue
Context Window272K tokens
Input Cost$30.00 / 1M tokens
Output Cost$180.00 / 1M tokens
HighlightsMaximum reasoning power; premium tier for hardest tasks
ReleasedMarch 2026

Best for: Competition-level math, complex multi-step reasoning. Considerations: Very expensive. GPT-5.4 Pro is still listed at $30 / $180, the same as GPT-5.5 Pro, the newer dedicated Pro ID. On the GPT-5.6 and GPT-6 families Pro is a mode (reasoning.mode: "pro") billed at the selected model's standard rates.

GPT-5.6 Sol / Terra / Luna (OpenAI) - GA July 9, 2026 (superseded by GPT-6 and GPT-6.1)

AttributeValue
VariantsSol (flagship), Terra (balanced), Luna (fast, low cost)
Context Window1.05M tokens (all three); 128K max output; knowledge cutoff February 16, 2026
Sol pricingPromotional $4 / $20 per 1M since August 21, 2026; OpenAI's pricing and model pages say the promotion is "available until November 30, 2026" and state no price after it. The prior $5 / $30 list is inferred from the stated 20% and 33% cuts, so budget on $5 / $30 from December 1
Terra pricing$2.00 / $12.00 per 1M (cut 20% from $2.50/$15 on July 30)
Luna pricing$0.20 / $1.20 per 1M (cut 80% from $1/$6 on July 30)
Long contextAbove 272K input the whole request bills at Sol $8 / $30 (promotional), Terra $4 / $18, Luna $0.40 / $1.80
Speed tiersFast (renamed from Priority on July 30) at 2x. Sol had an Ultrafast preview; OpenAI's current Ultrafast table lists only GPT-6 Astra and GPT-6.1 Sol
Reasoning"max" reasoning effort plus an "ultra" mode that uses subagents to accelerate complex work
API features at GAProgrammatic tool calling, multi-agent support, explicit prompt-cache breakpoints
BenchmarksSol: 53.6 on Agents' Last Exam (13.1 points ahead of Claude Fable 5) and a Terminal-Bench 2.1 record. Claude Fable 5 led SWE-Bench Pro (80 vs Sol's 64.6), a benchmark OpenAI publicly disputes. Vendor-reported. On the Terminal-Bench 4.0 leaderboard, Sol scores 37.27% at max
ReleasedLimited preview June 26, 2026; general availability July 9, 2026

What it is: OpenAI's first flagship line shipped as three models. The June 26 preview was gated at the request of the US government over dual-use cybersecurity capability; GA followed a 13-day review. The three-tier structure plus the July 30 cuts reset routing math for tiered model selection.

Status (November 2026): GPT-6 Sol and Luna replace Sol and Luna at about half the price, and GPT-6.1 Terra (October 26, $1/$5) replaces Terra at half its price. This family will stay in production for a while anyway: Foundry moves its retiring o-series deployments to GPT-5.6 Sol and Terra on November 19, and OpenAI's December 11 retirement of the dated GPT-5 and o3 snapshots points to the GPT-5.6 models.

GPT-5.6-Cyber (OpenAI) - August 2026 (restricted)

AttributeValue
Model IDgpt-5.6-cyber
Context Window400K total (272K max input, 128K max output)
Input / Output Cost$12.50 / $75.00 per 1M; cached input $1.25
AccessDaybreak Red tier only, with identity verification, legal attestations, and approved use cases. Responses API only. Requests take access_programs.cyber: "daybreak_red", and every Daybreak Red organization must require SSO for its API organization and enforce MFA through its identity provider
Refusal postureTrained for a lower refusal rate on dual-use security work: 95.0% completion on OpenAI's internal Advanced Cybersecurity Completion Rate eval versus 1.5% for GPT-5.6 Sol
ReleasedAugust 10, 2026 (the Daybreak Blue and Red tier split appears in the API changelog dated August 7)

What it is: A cybersecurity model built on GPT-5.6 Sol, shipped alongside a split of the Daybreak program into two tiers. Daybreak Blue gives approved defenders access to general-purpose frontier models for vulnerability discovery, secure code review, detection engineering, and incident response. Daybreak Red is separately approved and gates gpt-5.6-cyber for vulnerability reproduction, exploit validation, penetration testing, and red teaming.

Why it matters architecturally: A model that is deliberately more permissive than the frontier default, priced at 2.5x Sol's list price, restricted to a single API surface, and fenced behind identity verification is a reference design for how labs operationalize dual-use access. Gated models also retire fast: OpenAI deprecated the earlier gpt-5.4-cyber on September 11 and shut it down October 1, about 20 days' notice against its stated 3-month minimum for specialized variants.

GPT-6-Cyber (OpenAI) - November 2026 NEW (restricted)

AttributeValue
Model IDgpt-6-cyber
Context Window1,050,000 tokens (922,000 max input, 128,000 max output)
Input / Output Cost$20.00 / $100.00 per 1M; cached input $2.00; cache write $25 (30-minute) or $50 (12-hour). No long-context, Flex, or Fast rates; no Batch
APIChat Completions and Responses only
AccessDaybreak Red approval plus a separate model approval, access_programs.cyber: "daybreak_red", and IP allowlisting; rolling out to a select group first, widening over the following weeks
BenchmarksOpenAI-reported: outperforms GPT-6 Astra on most cyber tasks; used to find 81 previously unknown high-severity Firefox vulnerabilities and to develop working exploits for 948 Linux kernel vulnerabilities, 588 of them previously unknown
ReleasedNovember 5, 2026

What it is: the new top of the gated price ladder, at twice GPT-6 Astra's list price. The integration design changes more than the price: network egress (IP allowlisting) and identity (SSO and MFA at the organization level) become part of the contract, so plan a dedicated egress path and a separate API organization for it.

Gated access is now a request parameter. OpenAI's Daybreak guide documents a per-request access_programs.cyber field with values standard, daybreak_blue, and daybreak_red. GPT-5.6-Cyber and GPT-6-Cyber take daybreak_red; GPT-5.5-Cyber, GPT-6 Astra, and GPT-6.1 Sol take daybreak_blue but need Red approval for reduced refusals. The default changed with gpt-6.1-terra: from that model on, omitting the field gives standard treatment, while on older mainline models omission selects Blue if available (and daybreak_red on Red models). ZDR projects must use Private Safety Processing, and misconfigured requests fail with invalid_access_program (400), unsupported_access_program (400), or access_program_not_enabled (403). Set the field explicitly at the gateway on every route, so a model swap cannot silently change refusal behavior.

Capability-Tiered Access Programs (November 2026)

LabProgramGated model(s)Who gets in
AnthropicCyber Verification Program, expanded October 6 to three tiers (Defense, Red Team, Specialized) with Project Glasswing folded in; Life Sciences Verification ProgramClaude Mythos 5.1 (same weights as Fable 5.1, looser safeguards), available in every CVP tierVerified security organizations, with data retention required; life-sciences participants enrolled with the US government
OpenAIDaybreak Blue / Daybreak Red (selected per request with access_programs.cyber)Blue: general frontier models such as GPT-6 Astra and GPT-6.1 Sol, plus gpt-5.5-cyber (Daybreak organizations got Astra first); Red: gpt-5.6-cyber and gpt-6-cyber (November 5, $20 / $100; separate model approval and IP allowlisting)Approved defenders; Red is separately approved and requires SSO and MFA for the API organization
OpenAITrusted accessgpt-rosalind-research (life sciences; GA September 8 at $5 / $25 per 1M) and gpt-6-rosalind (October 30; $15 / $75, cached $1.50; $30 / $112.50 above 272K)Approved life-sciences organizations
GoogleFairwind ProgramGemini 3.8 Flash Cyber; Gemini 4 Carbon (announced October 28, a subset of cyber defenders, no model ID). Fairwind also got Gemini 4 Argon first, before its October 13 GAGovernment authorities, critical infrastructure operators, software maintainers

Design implication: capability-tiered access is now an industry pattern, not a one-off, and the gated tier is increasingly where each lab's strongest model lives first: Carbon tops the AA index while no developer can call it, GPT-6 Rosalind costs 7.5x GPT-6.1 Sol per input token, and GPT-6-Cyber 10x. Treat a gated tier as a separate model with its own identity requirements, API surface, retention terms, and a much shorter retirement runway.

GPT-5.5 (OpenAI) - April 2026

AttributeValue
Context Window1M tokens
Input Cost$5.00 / 1M tokens (cached $0.50)
Output Cost$30.00 / 1M tokens
Long context / FastAbove 272K input: $10 / $45. Fast is 2.5x ($12.50 / $75), not 2x as on newer models
MultimodalText, Image, Audio, Video
ARC-AGI-285.0% (leader as of May 13, 2026)
ReleasedApril 23, 2026

Best for: Multimodal workloads already validated on it. Pitched at launch as a "new class of intelligence for real work", it replaced GPT-5.4 for top-tier reasoning and multimodal work. Considerations: ~2× the input cost of GPT-5.4 ($2.50 → $5.00) and ~2× output ($15 → $30). For new builds, GPT-6 Sol ($2/$10) is cheaper and newer.

GPT-5.5 Instant (OpenAI) - May 2026

AttributeValue
StatusBecame the ChatGPT default and chat-latest in the API on May 5, 2026
Hallucination Reduction52.5% fewer on high-stakes prompts (medicine/law/finance) vs GPT-5.3 Instant
AIME 202581.2% (up from 65.4% on GPT-5.3 Instant)
Response Length~30% fewer words/lines than predecessor
ReleasedMay 5, 2026

Best for: ChatGPT-equivalent chat workloads, high-stakes domains where hallucination reduction matters. Considerations: Replaced GPT-5.3 Instant as the chat default. GPT-5.2-chat-latest and GPT-5.3-chat-latest deprecated May 8, 2026.

OpenAI Voice Models: GPT-Live-1, GPT-Realtime-2.1, Transcription (November 2026)

ModelShapePricingStatus
gpt-live-1Full-duplex voice front end that delegates reasoning and tools to a backend (an OpenAI Responses model or your own) on v1/live/sessions$0.05 per session minute, billed per second, plus backend model tokensGA September 10, 2026
gpt-realtime-2.1 / gpt-realtime-2.1-miniSingle speech-to-speech model on the Realtime API; turn-based with server_vad or semantic_vad; configurable reasoning effort2.1: audio $32 in / $0.40 cached / $64 out per 1M; mini: $10 / $0.30 / $20July 6, 2026 (succeeds gpt-realtime-2, May 7)
gpt-realtime-translateStreaming translation, one session per target language$0.034 per minuteCurrent
gpt-transcribe / gpt-live-transcribeBatch / streaming speech-to-text$0.0045 / $0.017 per minuteJuly 28, 2026
gpt-realtime, gpt-realtime-mini, gpt-4o-realtime, gpt-audioLegacy speech-to-speech and audio-Shut down January 20, 2027 (to gpt-realtime-2.1, -2.1-mini, or gpt-audio-1.5)
whisper-1, gpt-4o-transcribe familyLegacy speech-to-text-Shut down February 26, 2027 (to gpt-transcribe or gpt-live-transcribe)
tts-1, tts-1-hd, gpt-4o-mini-tts snapshotsLegacy text-to-speech-Shut down January 6, 2027 (announced October 1; to gpt-realtime-2.1-mini)

Three architectures, not two. OpenAI's docs now describe full-duplex-plus-delegation (GPT-Live), a single speech-to-speech model (Realtime API), and a chained STT, LLM, TTS pipeline. GPT-Live separates the voice layer from the reasoning layer, so business rules and authoritative task state belong in the backend. Once its 128K context passes 90% full, it restarts the voice engine with the instructions plus at most 8,192 tokens of history (including a summary of older turns), so keep slots and bookings in application state. Realtime sessions cap at 60 minutes and re-send the conversation on every response, so cost grows with call length. OpenAI reports GPT-Live turn-taking latency of 0.798 s vs 1.41 s for gpt-realtime-2.1 (vendor-reported, and dependent on the backend you pick). The Realtime API beta interface was removed May 12, 2026.

Other vendors: Google's gemini-3.8-live and gemini-3.8-live-extended-thinking went GA September 15 (audio $3 in / $12 out per 1M; asynchronous tool calls by default, and the Extended Thinking variant reasons in the background while it talks). Google reports tau-Voice 68.6% for Extended Thinking against 30.1% for the base model (vendor-reported), which is the pattern to design for: voice task success tracks the reasoning attached to the voice layer, and that reasoning has to fit the turn latency budget. gemini-3.8-flash-tts went GA September 22 at a promotional $0.50 / $9 per 1M through December 31 ($1 / $18 from January 1, 2027). Voice stacks still on Google preview IDs face hard shutdowns: on November 17 (scheduled, an earliest-possible date with no routing described) the Gemini API shuts down gemini-3.1-flash-live-preview and gemini-2.5-flash-native-audio-preview-12-2025 (replacement gemini-3.8-live) and the TTS previews gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, and gemini-3.1-flash-tts-preview (replacement gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts); on Vertex AI, gemini-live-2.5-flash-native-audio retires December 13 (to gemini-3.8-live). xAI bills voice at $0.08 per minute, and in October its grok-voice-latest alias moved from grok-voice-2.0 to grok-voice-3.0 at the same price: an alias upgrade that changes the model under unchanged code. Anthropic has no speech API: Claude can only be the text backend in a voice stack, for example behind GPT-Live client delegation. See Realtime Voice Agents.

Gemini 4 Argon (Google) - October 2026 NEW (GA October 13)

AttributeValue
Model IDgemini-4-argon (stable; paid tier only, not on the free tier)
Context Window1,048,576 tokens shared between input and output; output defaults to 65,536 and reaches about 1M only through "long decode", a continuation protocol
Input / Output Cost$2 / $10 per 1M through November 30, 2026; $4 / $20 from December 1. No prompt-length tiers
CacheImplicit caching only (hits $0.10, then $0.20 from December 1); explicit caching is not supported
Service tiersBatch and Flex half of Standard; Priority $3.60 / $18 ($7.20 / $36 from December 1)
Modalities and toolsText, image, video, audio, and PDF in; text out. Function calling, structured outputs, computer use, Search grounding, Batch; no Live API
SafetyModel card: Google "cannot rule out" that it reaches the critical capability levels for biology and cybersecurity
BenchmarksGoogle-reported at the September 30 announcement: DeepSWE v1.1 77.9%, AutomationBench 51.3%, LVBench 91.7%. Independent: AA Intelligence Index v5.0 50 (high)
ReleasedAnnounced September 30, 2026 (Fairwind cyber defenders first); GA October 13, 2026

How to treat it: budget at $4/$20, because the introductory price lasts seven weeks. At that price it matches Claude Opus 5.5's list price while scoring 8 points lower on AA v5.0 (50 vs 58), so its case is native video and audio input, Google's tooling, and long output, not raw capability per dollar. Two contract details change architecture. The 1M output only arrives through the long-decode protocol, so a whole-repo rewrite still needs a continuation loop, just a vendor-managed one. And with no explicit caching, cache-control code written for earlier Gemini models does nothing; cached cost depends on implicit hit rates you cannot pin, so measure them before trusting a cached-cost estimate. Argon is a new Gemini 4 line, not the Gemini 3.5 Pro that never shipped; gemini-3.1-pro-preview stays listed with no shutdown date.

Gemini 4 Carbon (Google) - ANNOUNCED October 28, 2026, not available to developers

Google announced Carbon on October 28 for a subset of cyber defenders through the Fairwind Program. There is no model ID and no date for broader access. Artificial Analysis added it on October 28 at 59 (high), the top score on Intelligence Index v5.0, ahead of Claude Opus 5.5 (58); after the November 3 v5.0.1 regrade it scores 60 and Opus 5.5 59. Treat it like Argon in September: a signal of where the frontier is, not a dependency. If a leaderboard's top entry is a model you cannot call, compare against the best model you can.

Gemini 3.8 Flash (Google) - September 2026

AttributeValue
Model IDgemini-3.8-flash (stable, GA)
Context Window1,048,576 tokens in / 65,536 out
Input / Output Cost$0.75 / $3.75 per 1M through December 31, 2026; $1.50 / $7.50 from January 1, 2027. Batch and Flex 50% off; Priority $1.35 / $6.75 ($2.70 / $13.50 from January 1); cache read $0.075 per 1M plus storage
ModalitiesText, image, video, audio, PDF in; text out
Thinkingthinking_level low, medium (default), high; no minimal level; replaces thinking_budget
ToolsFunction calling, structured outputs, caching, code execution, computer use (the Gemini Computer Use tool went GA October 30), file search, Search and Maps grounding; no Live API
BenchmarksGoogle-reported: DeepSWE v1.1 73.7% vs Claude Opus 5 74.0%. Independent: AA Intelligence Index v5.0 31 (high); SWE-Bench Pro v2 private set 211/272 (Scale)
ReleasedSeptember 2, 2026

What it is: Further training on Gemini 3.7 Flash rather than a new base, and Google's third Flash in six weeks. Since October 8, requests to gemini-3.7-flash are served by 3.8 Flash automatically. It is the Gemini model Google recommends for computer use. The per-token price is low, but coverage notes it spends more reasoning tokens per task. On DeepSWE v1.1, where it tied Astra and Opus 5 near 74%, its cost per task was $2.36 at high effort versus $4.43 for Astra at xhigh and $11.84 for Opus 5 at max. Model the January 2027 price before committing volume: 3.6, 3.7, and 3.8 Flash all double on the same day.

Gemini 3.8 Flash Cyber: announced the same day for trusted defenders in the new Fairwind Program, with more permissive cybersecurity mitigations and Google's CodeMender harness for find-verify-fix pipelines (Google reports CyberGym 86.2%). It replaces Gemini 3.5 Flash Cyber in Google's tiered-access lineup.

Gemini 3.1 Pro Preview (Google) - February 2026

AttributeValue
Model IDgemini-3.1-pro-preview (Preview)
Context Window1M tokens
Input Cost$2.00 / 1M tokens (standard); $4.00 (200K+)
Output Cost$12.00 / 1M tokens (standard); $18.00 (200K+)
MultimodalNative: Text, Vision, Audio, Video
HighlightsStrong agentic and coding capabilities; still a preview, with Gemini 4 Argon now the GA model above it
ReleasedFebruary 2026

Best for: Complex reasoning, multimodal analysis, long-context workloads. Considerations: Replaced Gemini 3 Pro Preview. It is still the only 3-series Pro model on the Gemini API (there is no Gemini 3.5 Pro), and it has no shutdown date; for new work at this price point, compare it with Gemini 4 Argon at Argon's December price. Gemini 2.5 is now deprecated on the Gemini API. After restricting the 2.5 models to existing users on September 18 and calling them "not deprecated", Google deprecated gemini-2.5-pro and gemini-2.5-flash on November 2, 2026; both are no longer selectable in Google AI Studio. From December 1 (scheduled), gemini-2.5-pro requests route to gemini-3.8-flash and bill at 3.8 Flash rates, and gemini-2.5-flash requests route to gemini-3.5-flash-lite at its rates. Pinned 2.5 Pro code will keep returning 200s while a different model answers, with breaking changes Google lists on the targets: stricter function-calling and thought_signature validation, thinking_level replacing thinking_budget, removed sampling parameters, and no prefilled model turns. gemini-2.5-flash-lite is not deprecated. Google Cloud (Vertex AI) scheduled gemini-2.5-pro, gemini-2.5-flash, and gemini-2.5-flash-lite to retire on October 20, 2026; as of November 6 that retirement is still not confirmed as executed, so test before assuming either outcome.

Gemini Flash and Flash-Lite Lineup (Google)

ModelInput / Output per 1MStatus
gemini-3.8-flash$0.75 / $3.75 intro; $1.50 / $7.50 from January 1, 2027GA September 2, 2026; the current default
gemini-3.7-flashSame intro pricing as 3.8 FlashRequests route automatically to gemini-3.8-flash since October 8. On Vertex AI it retires January 28, 2027 (scheduled), with no routing described
gemini-3.6-flashSame intro pricing as 3.8 FlashSuperseded by 3.8 Flash; now also the target for gemini-3.5-flash traffic. On Vertex AI it retires November 19, 2026 (scheduled, replacement gemini-3.8-flash, no auto-upgrade), while Vertex keeps gemini-3.5-flash until "May 19, 2027 or later": the same ID has a different lifecycle per platform
Gemini 3.5 Flash$1.50 / $9.00Requests route automatically to gemini-3.6-flash since October 8; Google has not said how rerouted requests bill
Gemini 3.5 Flash-Lite$0.30 / $2.50Google's pointer, with 3.8 Flash, for projects leaving 2.5
Gemini 3.1 Flash-Lite$0.25 / $1.50Shuts down May 7, 2027 (to 3.5 Flash-Lite)
Gemini 3 Flash Preview$0.50 / $3.00Preview
Gemini 2.5 Flash-Lite$0.10 / $0.40Existing users only since September 18; not deprecated, unlike 2.5 Pro and 2.5 Flash (November 2)

Watch for a phantom row: a "Gemini 3.1 Flash" at $0.10 / $3.00 still circulates in older cost tables. No model with that name and price appears on Google's pricing page; the nearest are Gemini 3 Flash Preview and Gemini 3.1 Flash-Lite. Price new work on Gemini 3.8 Flash.

Gemini 3.2 Flash (Google) - May 2026, Historical

AttributeValue
StatusQuiet rollout in iOS Gemini app and Google AI Studio May 5, 2026 (no formal announcement)
ReleasedMay 5, 2026

Status: Historical. Google's Flash line has since moved through 3.5, 3.6, 3.7, and 3.8; build on gemini-3.8-flash.

Gemini Deep Research / Deep Research Max (Google) - April 2026

AttributeValue
Built onGemini 3.1 Pro
CapabilitiesMCP support; native chart/infographic generation; extended test-time compute; async background workflows
ReleasedApril 21, 2026

Best for: Research agents, document synthesis, long-running async workflows. The MCP support made it the first Google research-agent product with first-class tool integration.

Gemini Robotics-ER 1.6 (Google DeepMind) - April 2026

AttributeValue
DomainPhysical robotics, embodied reasoning
New capabilityReading gauges/sight glasses
DeploymentBoston Dynamics Spot
ReleasedApril 14, 2026

Best for: Robotics applications requiring vision-language grounding for physical actions. Available via Gemini API and AI Studio.

Gemini 3.7 Flash (Google) - August 2026, SUPERSEDED by 3.8 Flash (rerouted October 8)

AttributeValue
Model IDgemini-3.7-flash
Context Window1,048,576 tokens in / 65,536 out
Input / Output Cost$0.75 / $3.75 per 1M through December 31, 2026, then $1.50 / $7.50. Context caching $0.075 per 1M; Batch $0.375 / $1.875
MultimodalText, image, audio, video; agentic video understanding is opt-in ("processing": "agentic"), not on by default
Knowledge cutoffMarch 2026
ReleasedAugust 13, 2026 (GA)

What it is: Google's workhorse tier, built on Gemini 3.6 Flash with what the model card describes as algorithmic improvements to the reasoning foundation rather than a new pretraining run. Customizable thinking configurations trade quality against cost and latency per request. Google-reported gains over 3.6 Flash include FrontierCode 1.1 Main 43.6% versus 34.4% and WebDev Arena Elo 1588 versus 1538.

Agentic video (September 1): with "processing": "agentic" on a video input, the model decides which segments to inspect, at what speed, and through which modality, instead of sampling at a fixed 1 FPS (still the default). Google reports up to 88% fewer tokens and up to 66% lower cost at standard token pricing, and the mode now covers 3.8 Flash, 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Video cost models built on fixed-FPS sampling can overestimate tokens by as much as roughly 8x.

Considerations: Superseded by 3.8 Flash on September 2 at the same intro price, and since October 8 requests to gemini-3.7-flash route to gemini-3.8-flash automatically: a pinned model string kept working while the model behind it changed, so evals pinned to 3.7 Flash now measure 3.8 Flash. Gemini 3.5 Pro never shipped; Google's next flagship was Gemini 4 Argon (GA October 13).

Grok 4.7 (SpaceXAI) - September 2026

AttributeValue
Model IDgrok-4.7
Context Window500K tokens; text and image in, text out
Input / Output Cost$2.00 / $6.00 per 1M below a 200K prompt; $4.00 / $12.00 on all tokens once the prompt reaches 200K. Cached input $0.50 ($1.00 at 200K+)
Fast variantAbout 2x output speed at 2x price, offered only in Cursor and Grok Build (not a listed API model)
Knowledge cutoffMay 2026 per the API docs
BenchmarksTerminal-Bench 4.0 leaderboard 37.58% at xhigh. AA Intelligence Index v5.0: 35 (high)
ReleasedSeptember 21, 2026 (Grok API, Grok Build, Cursor on all plans, OpenRouter and other gateways)

What it is: Unlike most August releases, a new and larger base model, with a longer RL run weighted toward multi-hour tasks (per SpaceXAI). It is also this quarter's clearest case of vendor tables outrunning independent indices: large gains in SpaceXAI's own table, small ones on AA, where v5.0 puts it about 20 points behind GPT-6.1 Sol (55) and Sonnet 5.5 (54) at the same $2 input price. Still SpaceXAI's newest text model: there is no Grok 5 as of November 6. Distribution now runs through Cursor, which became part of SpaceX in mid-August, and Grok launches appear on Cursor's blog. The whole-request 200K pricing trap carries over from 4.6.

Grok 4.6 (SpaceXAI) - August 2026, SUPERSEDED by Grok 4.7

AttributeValue
Model IDgrok-4.6
Context Window500K tokens
Input / Output Cost$2.00 / $6.00 per 1M below a 200K prompt; $4.00 / $12.00 at or above 200K. Cached input $0.50
ReasoningEffort settings low, medium, high (default), xhigh
Knowledge cutoffFebruary 1, 2026
ReleasedAugust 12, 2026

What it is: SpaceXAI's frontier model before 4.7, focused on long-running agents and interactive visual work. It scored 61 on the Artificial Analysis Intelligence Index at launch, up from 56 for Grok 4.5 High, on the pre-September index. AA has re-baselined twice since (v4.3 in September, v5.0 in October), so that number compares with nothing current.

Two traps worth knowing. The higher long-prompt rate applies to all tokens in the request once the prompt reaches 200K, not just the tokens past the threshold. And cached input got more expensive than Grok 4.5 ($0.50 versus $0.30 per 1M), so cache-heavy agent loops do not automatically get cheaper on the upgrade. Note the vendor name: xAI completed a rebrand to SpaceXAI, so current docs and release notes use the new name.

Grok 4 (xAI) - RETIRED on the xAI API

AttributeValue
Context Window256K tokens
Input Cost$3.00 / 1M tokens
Output Cost$15.00 / 1M tokens
HighlightsNative tool use and real-time search; competitive reasoning
ReleasedJuly 2025 (Grok 4.20 beta: February 2026)

Status: Retired from the xAI API on May 15, 2026, along with grok-4-fast, Grok 4.1 Fast, grok-code-fast-1, and grok-3. The retired text-model slugs still resolve but redirect to grok-4.3 ($1.25 / $2.50) and bill at its rates (grok-code-fast-1 goes to grok-build-0.1), so code keeps "working" while behavior and cost change. Azure Foundry still lists grok-4 and grok-4-1-fast as GA, a reminder that lifecycle is tracked per (model, platform) pair.

MAI-Thinking-1 (Microsoft) - August 2026 (public preview)

AttributeValue
StatusPublic preview on Microsoft Foundry since August 12, 2026
ArchitectureSparse MoE, about 1T total / 35B active parameters; 256K context
APIChat Completions compatible, function calling
PricingFrom $2 input / $8 output per 1M (Microsoft's Foundry blog, updated August 12)
LimitsPreview version 2026-06-01; Global Standard only, no PTU; 256K total tokens per request with output capped at 64K; function tools only; an opaque encrypted reasoning state when reasoning_display is encrypted; default quota 0 TPM and 0 RPM (Medium 100,000 TPM / 100 RPM, High 250,000 TPM / 250 RPM)
BenchmarksMicrosoft-reported: AIME 2025 97.0%, SWE-Bench Pro on par with Claude Opus 4.6, blind human preference over Claude Sonnet 4.6 across 1,276 tasks

Why it matters: Microsoft now has a first-party reasoning model in Foundry next to its OpenAI and Anthropic offerings, which matters for vendor diversification and sovereignty discussions. Microsoft says it did not distill from other labs' models. Price is no longer the blocker: it lists at GPT-6 Sol's input price and below it on output. The real blockers are the zero default quota (request capacity before you design around it) and the 64K output cap.

Model Comparison: Frontier and Workhorse Tiers (November 2026)

ModelInput / Output per 1MCache readContextAccessStatus
Claude Fable 5.1$10 / $50$0.25 (0.025x)1M, flat pricing5 platforms; ZDR only if authorizedGA September 1
GPT-6 Astra$10 / $50 ($20 / $75 above 272K)$1.001.05MAPI, Bedrock, Foundry; Google Cloud (private preview)Staged GA from September 3
Claude Opus 5.5$4 / $20$0.20 (0.05x)1M, flat pricing5 platforms; ZDRGA September 22
Meta Muse Wonder$4 / $20$0.201MMeta Model API, Muse Code (Muse Wonder Full: Muse Code subscription only)Released October 12
Gemini 4 Argon$2 / $10 through November 30; $4 / $20 from December 1$0.10 implicit only ($0.20 from December 1)1M shared in and out; 65,536 default outputGemini API paid tierGA October 13
Claude Sonnet 5.5$2 / $10$0.10 (0.05x)1M, flat pricing5 platforms; ZDRGA September 28
GPT-6.1 Sol$2 / $10 ($4 / $15 above 272K)$0.10 (0.05x)1.05MAPI, Bedrock; Google Cloud (private preview)GA September 29
GPT-6 Sol$2 / $10 ($4 / $15 above 272K)$0.201.05MAPI, BedrockGA September 22
Grok 4.7$2 / $6 ($4 / $12 at 200K+)$0.50500KxAI API, Cursor, gatewaysGA September 21
Muse Spark 1.3$1.25 / $4.25$0.151MMeta Model API, Muse CodeGA September 2
GPT-6.1 Terra$1 / $5 ($2 / $7.50 above 272K)$0.05 (0.05x)1.05MAPI, BedrockGA October 26
Gemini 3.8 Flash$0.75 / $3.75 intro ($1.50 / $7.50 from January 1)$0.0751M in / 65,536 outGemini API, Gemini EnterpriseGA September 2
Claude Haiku 5.5$0.10 / $0.50 up to a 100K prompt; $0.50 / $2.50 on the whole request above$0.01 ($0.05 above 100K)1M5 platformsGA October 7
GPT-6 Luna$0.10 / $0.50 ($0.20 / $0.75 above 272K)$0.011.05MAPI, Bedrock; Google Cloud (private preview)GA September 22

"5 platforms" = Claude API, Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS. Since October 22 Google Cloud offers Claude two ways, Google managed and Anthropic managed (Claude Models on Google Cloud, which follows Anthropic's lifecycle dates), so Anthropic now counts six platforms. Standard tier, list prices. Benchmarks for these models are in Capability Comparison; full pricing mechanics (batch, speed tiers, residency premiums) are in Pricing and Costs.

Production Heritage & Maturity

While frontier models lead on benchmarks, many enterprise systems rely on battle-tested models, and most of them now have a retirement date:

Model FamilyProduction SinceMaturity Note
GPT-4o / GPT-4.1May 2024 / April 2025Most mature ecosystem, now on a short clock. The original gpt-4o-2024-05-13 snapshot shut down on the OpenAI API on October 23, 2026. On October 14 OpenAI gave about three months' notice that gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, and fine-tunes of them shut down January 14, 2027, replaced by gpt-6-luna, which cannot be fine-tuned. On Microsoft Foundry, gpt-4o 2024-05-13 retires the same day and its Global Standard and Data Zone Standard deployments auto-upgrade to gpt-5.1 (2025-11-13), a reasoning-class model swapped in with no code change; the other GPT-4o and GPT-4.1 versions run on Foundry until April 14, 2027.
Claude Sonnet 4.5 / 4.6Sept 2025 / Feb 2026Inherited the tool-use reliability reputation of Claude 3.5 and 3.7 Sonnet. Sonnet 4.5 retires November 30, 2026 on the Claude API and Foundry but runs on Bedrock until April 8, 2027; Sonnet 4.6 is legacy and now costs more than Sonnet 5.5.
Gemini 2.5 ProMarch 2025Proven long-context. Restricted to existing users on the Gemini API since September 18, 2026, then deprecated there on November 2; from December 1 its requests route to gemini-3.8-flash at 3.8 Flash rates. Google Cloud (Vertex) scheduled its retirement for October 20, 2026; execution is not confirmed.
o1 / o3Sept 2024Well-understood reasoning failure modes. The OpenAI API shut down o1, o1-pro, o3-mini, and o4-mini on October 23; Foundry retires o1, o3, o3-pro, and siblings on November 19 (to GPT-5.6 Sol); OpenAI shuts down o3-2025-04-16 and the dated GPT-5 snapshots on December 11.

Why stay on "older" frontier models?

  1. Consistency: New models have release-window latency spikes and behavior shifts. GPT-6 Sol's image path was degraded until a fix three days after launch.
  2. Cost is no longer the reason: Below the ceiling tier (where GPT-6 Astra costs more than GPT-5.6 Sol), the new generation is cheaper. Opus 5.5 undercuts Opus 5 by 20%, GPT-6 Sol and Luna cost roughly half of GPT-5.6 Sol and Luna, GPT-6.1 Terra half of GPT-5.6 Terra, Haiku 5.5 a tenth of Haiku 4.5 below 100K tokens, and Sonnet 5.5 less than Sonnet 4.6. Staying put now costs money.
  3. Guardrail Tuning: Security and moderation layers are more refined.
  4. The runway is short, and notice policies are not guarantees: OpenAI's stated minimum is 6 months for GA models, 3 months for specialized variants, and as little as 2 weeks for previews, yet its October 14 GPT-4o and GPT-4.1 notice gave about 3 months; Anthropic gives at least 60 days; new Bedrock models get 6 months or 45 days, but Bedrock gave several Claude 3.x regions about a month of Legacy before a November 13 end of life. Pin dated IDs and track retirements per (model, platform) pair.

Retirement Calendar (as of November 6, 2026)

DateWhat changesMove to
November 13Bedrock Claude 3 Haiku, 3 Sonnet, 3.5 Sonnet, 3.5 Sonnet v2, and 3.7 Sonnet in the remaining listed regions (Legacy only since October 10 to 15)A current Claude model on Bedrock
November 15Foundry claude-haiku-4-5 (Anthropic's own floor implies about January 2027; plan for the earlier date) and codex-miniClaude Haiku 5.5
November 17Gemini API preview speech models: gemini-3.1-flash-live-preview, gemini-2.5-flash-native-audio-preview-12-2025, gemini-2.5-flash-preview-tts, gemini-2.5-pro-preview-tts, gemini-3.1-flash-tts-preview (earliest-possible date, no routing)gemini-3.8-live; gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts
November 19Foundry o1, o1-pro, o3, o3-pro, o3-deep-research; o3-mini, o4-mini. Vertex AI gemini-3.6-flash (no auto-upgrade)GPT-5.6 Sol; GPT-5.6 Terra; gemini-3.8-flash
November 24Foundry claude-opus-4-5. The Claude API floor passes the same day with no notice issued, so not before about January 2027 thereOpus 5.5
November 26Bedrock AI21 Jamba 1.5-
November 27Bedrock Claude 3 Haiku and Claude 3 Sonnet in ca-central-1 (Legacy since October 15), a second cutoff the November 13 row hidesA current Claude model on Bedrock
November 30claude-sonnet-4-5-20250929 (Claude API, Foundry); Opus 4.8 fast mode removed (speed: "fast" errors); OpenAI Evals, Agent Builder, and Reusable Prompts; last day of GPT-5.6 Sol's promotional $4 / $20claude-sonnet-5-5; Opus 5.5 fast mode; Promptfoo (OpenAI-owned, MIT) for evals; budget GPT-5.6 Sol at list or move to GPT-6.1 Sol
December 1Gemini 4 Argon doubles to $4 / $20 (cache hits $0.20); gemini-2.5-pro and gemini-2.5-flash requests start routing to gemini-3.8-flash and gemini-3.5-flash-lite at those models' rates; OpenAI gpt-image-1.5, gpt-image-1-mini, chatgpt-image-latest retireBudget Argon at $4 / $20; migrate 2.5 code deliberately and test tool calls; gpt-image-2.5-sunburst or -flare, or the same-priced gpt-image-2.6-sunburst or -flare (November 4; no Batch rate listed yet)
December 11gpt-5-2025-08-07, o3-2025-04-16, and sibling snapshotsGPT-5.6 Sol, Terra, or Luna
December 13Vertex AI gemini-live-2.5-flash-native-audiogemini-3.8-live
December 15Foundry whisper, tts, tts-hdCurrent OpenAI speech models
End of DecemberMoonshot plans to retire all Kimi K2 series models on its API, including K2.6 and K2.7 Code (closed to newly registered users since October 31)Kimi K3.1 Pro
January 6, 2027tts-1, tts-1-hd, gpt-4o-mini-tts snapshots; OpenAI fine-tuning job creation endsgpt-realtime-2.1-mini for speech
January 8, 2027Bedrock Claude Opus 4.1 end of life; Bedrock Claude Sonnet 4.5 enters higher-priced public extended accessOpus 5.5; Sonnet 5.5
January 14, 2027OpenAI gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14, and fine-tunes of them; Foundry gpt-4o 2024-05-13 (Standard deployments auto-upgrade to gpt-5.1)gpt-6-luna (no fine-tuning; move tuned workloads to open weights or a cloud that still tunes)
January 20, 2027gpt-realtime, gpt-audio, and the gpt-4o realtime/audio familiesgpt-realtime-2.1, gpt-audio-1.5
January 28, 2027Vertex AI gemini-3.7-flash (no auto-upgrade described)gemini-3.8-flash
February 2, 2027Foundry claude-opus-4-6, three days before Anthropic's own floor (February 5)Opus 5.5
February 10, 2027Foundry claude-sonnet-4-6, a week before Anthropic's own floor (February 17)Sonnet 5.5
February 26, 2027whisper-1, gpt-4o-transcribe familygpt-transcribe, gpt-live-transcribe
March 15, 2027gemini-2.5-flash-image (moved from October 2)A current Gemini image model
March 30, 2027Bedrock Legacy batch: Llama 3.x and Llama 4 Scout and Maverick, DeepSeek-R1 and V3.1, Gemma 3, Qwen3-Coder, GLM 4.7, MiniMax M2 and M2.1, Kimi K2 Thinking, Devstral 2, Voxtral, Nemotron NanoA current open model on Bedrock, or self-host
April 1, 2027gpt-5.1, gpt-5.3-codex, gpt-5.4-nanoGPT-6 Sol or GPT-6 Luna
April 8, 2027Bedrock Claude Sonnet 4.5 end of lifeSonnet 5.5
April 14, 2027Foundry gpt-4o (2024-08-06, 2024-11-20), gpt-4o-mini, and the gpt-4.1 family, including gpt-4.1-nano (moved from October 14)A current OpenAI model on Foundry
April 30, 2027Mistral Medium 3.5, Mistral Large 3, and the Voxtral TTS 26.03 API model (all deprecated October 31)Mistral Large 4 (for Large 3 users, a 40% price increase); Voxtral TTS 3 (public preview)

Already past (October 2 to November 6):

  • Executed October 29 to November 6: Google deprecated gemini-2.5-pro and gemini-2.5-flash on the Gemini API (November 2; routing starts December 1). xAI retired grok-imagine-image-quality in November (scheduled for November 2; xAI's notes give only the month): requests to it and to grok-imagine-image-pro are now served by grok-imagine-image-2.0-low, the 2.0 model at low quality, at a lower per-image price with no request changes, so set quality explicitly if you need more than low. Moonshot closed the Kimi K2 series to newly registered API users (October 31), Mistral deprecated Voxtral TTS 26.03 (October 31), and Anthropic moved consumer Haiku 4.5 chats to Haiku 5.5 (October 29).
  • Executed: OpenAI shut down its legacy snapshots on October 23 as scheduled (gpt-4o-2024-05-13, gpt-4-turbo, o1, o3-mini, o4-mini, gpt-image-1, and the others announced April 22, plus their fine-tunes), and gpt-3.5-turbo-instruct, babbage-002, and davinci-002 on September 28. Bedrock lists Claude 3 Haiku (September 10, including GovCloud), Nova Sonic v1 (September 14), Nova Canvas and Nova Reel (September 30), and Claude Sonnet 4 (October 14) as reached end of life; Opus 4.1 entered public extended access October 8 and Sonnet 4.5 moved to Legacy the same day.
  • Behavior changes with no retirement: Anthropic API version 2026-10-21 (October 21) returns a 400 for temperature, top_p, and top_k at any value on every model, and the new SDK majors send it by default, so it lands whenever a team upgrades its SDK. Google rerouted gemini-3.7-flash to 3.8 Flash and gemini-3.5-flash to 3.6 Flash (October 8), and new paid Gemini projects store requests by default from October 26 (store=true).
  • Moved: gemini-2.5-flash-image to March 15, 2027; Foundry gpt-4.1-nano to April 14, 2027.
  • Not executed: the Claude Haiku 4.5 floor (October 15) passed with no deprecation notice; it is still Active on the Claude API as of November 6.
  • Not confirmed: Google Cloud's Vertex retirement of gemini-2.5-pro, -flash, and -flash-lite (October 20); Foundry gpt-4o-transcribe and gpt-4o-mini-transcribe 2025-03-20 (October 15; still listed as GA); OpenAI Evals going read-only (October 31; shutdown is still November 30); antigravity-preview-05-2026 (October 5); Mistral's GLM 5.2 (October 31) and OCR 4.0 and Leanstral 1.5 (September 30), past their listed dates with no extension; and Bedrock Nova Premier (September 14, in neither Bedrock's EOL table nor its catalog). Treat any of these still in your code as broken until you have tested them.
  • Earlier (August 15 to October 1): Groq's free and developer tiers dropped Llama 3.3 70B and Llama 3.1 8B (August 16), Anthropic Workbench closed (August 17), the OpenAI and Azure Assistants API shut down (August 26), Mistral Medium 3.1 retired (August 31), OpenAI Sora 2 and the Videos API shut down (September 24), and gpt-5.4-cyber shut down on 20 days' notice (October 1).

Open Source Models

Open-Weight Standings (November 2026)

Artificial Analysis Intelligence Index v5.0 (announced October 27): four equally weighted categories (Knowledge Work, Coding, Science, General), about 60% of it on private or held-out sets. On November 3 AA shipped methodology v5.0.1, which applied grading-robustness fixes from the CritPt authors and regraded every model on CritPt v2. v4.x scores are not comparable, so do not set these beside September or early-October figures. Scores marked v5.0.1 come from the current board; unmarked ones are v5.0 figures from before the regrade. Scores are the best listed effort for each model; a dash means AA shows no score.

ModelAA scoreTotal / activeLicense
Moonshot Kimi K3.1 Pro (weights November 4; AA still labels it proprietary)46 (v5.0.1)2.8T / 104BKimi K3.1-Pro License (MIT-like grant; MaaS agreement above US$20M revenue)
DeepSeek V4.1-Flash-101733 (v5.0.1, max); 29 at high, 19 non-reasoning552B backbone (763B checkpoint); 8B prefill / 16B decode activeMIT
Z.ai GLM-5.332 (v5.0.1, max); 23 at low744B / 40BGLM-5.3 License (MIT plus a MaaS security review above US$10B revenue)
Xiaomi MiMo-V2.6-Pro30 (v5.0.1)1.02T / 42BMIT
Ant Group Ling-3.1-flash-101930 (v5.0.1)560B / about 25BMIT
StepFun Step 5 Preview29600B / 27BStepFun Model License 1.0 (MaaS agreement above US$20M revenue; separate license for branded devices)
MiniMax-M3.1-Flash29About 428B / 23B per the model cardMiniMax Community License (commercial use only under conditions)
Z.ai GLM-5.3-Flash28320B / 18BMIT
Mistral Large 426 at High; 23 at MediumAbout 1T / 49B (52B with embeddings)Mistral Ecosystem Open-Weight License (separate license for any MaaS business)
Meta Muse Spark Open25 (v5.0.1)816B / 42BMuse Model License (MaaS agreement above US$20M revenue)
Alibaba Qwen3.8-27B20 (v5.0.1, xhigh)27B denseApache 2.0
BOSS Zhipin Nanbeige4.3-33B20 (v5.0.1)33B denseApache 2.0
Reflection AI Beam 1.018 (v5.0.1)501.3B / 23.4BApache 2.0
MBZUAI K2 Horizon 375B-A23B18 (v5.0.1)375B / 23BApache 2.0, open training data
NVIDIA Nemotron 3.5 Super VL14120B / 12BOpenMDW-1.1
Thinking Machines Inkling-Small13~266B / 12BApache 2.0
Thinking Machines Inkling12975B / 41BApache 2.0
NVIDIA Nemotron 3 Ultra12550B / 55BOpenMDW-1.1
Moonshot Kimi K3-2.8T / 104BKimi K3 License (MaaS agreement above US$20M revenue)
Alibaba Qwen3.8 2.4T-A95B-2.4T / 95BQwen3.8-Max License
Ricursive Ripple-65B Preview- (not yet scored)65B denseApache 2.0

For scale, the best model developers can call is Claude Opus 5.5 at 59 on v5.0.1, and the not-yet-public Gemini 4 Carbon scores 60. The open-to-closed gap now depends on the classification rule, so quote it with the index version and the rule. By weight availability, Kimi K3.1 Pro (weights published November 4) is the top open model at 46, 13 points behind Opus 5.5. By AA's own label, which still read "Proprietary" on November 6, the leader is DeepSeek V4.1 Flash 1017 at 33, 26 points behind. AA's October 27 v5 announcement had already counted K3.1 Pro as "the top open weights model at 46" and wrote that the open-weights gap widened from 3 points on v4.3 to 12, while the released-weight leaders sat 26 behind. Effort matters as much as the model: DeepSeek falls from 33 at max to 29 at high and 19 without reasoning, and GLM-5.3 from 32 at max to 23 at low, so evaluate at the effort you will pay for. Muse Spark Open (25) is still the top US open-weight model, ahead of Reflection's Beam 1.0 and MBZUAI's K2 Horizon 375B (both 18) and well ahead of both Inkling models.

Architecture note: the new open flagships all ship non-standard attention or KV schemes: sliding-window plus global layers on MiMo-V2.6-Pro, linear plus sparse attention on GLM-5.3-Flash and Qwen3.8-Flash-Next, sparse attention on Tencent Hy4, a hybrid linear MoE on Ling-3.1-flash (63 KDA linear-attention layers and 9 gated MLA layers), and a causal encoder-decoder with 890 bytes of global KV per token on DeepSeek V4.1-Flash. The textbook KV-size formula no longer transfers between models, so size capacity from each model card (see KV Cache and Context Caching). Speculative-decoding heads now ship inside checkpoints too: DSpark in DeepSeek V4-Pro-0813, MTP heads in MiMo-V2.6, Qwen3.8-Flash-Next, and Ling-3.1-flash, and a DFlash2 drafter alongside Reflection's Beam 1.0.

Llama 4 Family (Meta) - April 2025

ModelParametersContextArchitectureNotes
Llama 4 Scout17B active / 16 experts, 109B total (MoE)10MSparse MoEIndustry-leading 10M context; fits a single H100 at Int4; beat Gemma 3 and Gemini 2.0 Flash-Lite per Meta
Llama 4 Maverick17B active / 128 experts, 400B total (MoE)1MSparse MoEBeat GPT-4o and Gemini 2.0 Flash per Meta; comparable to DeepSeek V3 at half the active params
Llama 4 Behemoth~288B active, about 2T total-Sparse MoENever released; Meta paused it in 2026, and its frontier models now ship closed as Muse Spark

Strengths:

  • First Llama generation with Mixture-of-Experts architecture
  • Natively multimodal from the ground up (text, image, video input)
  • Open weights on Hugging Face under the Llama 4 Community License; available via Meta AI on WhatsApp, Messenger, Instagram
  • Scout's 10M token context window is industry-leading for open models

Watch out: there is no Llama 4 8B, 70B, or 405B. Dense worked examples should use Llama 3.3 70B or Llama 3.1 405B, or a current open model. Fast-inference hosts are also dropping Llama: Groq retired Llama 4 Scout (July 17) and Llama 3.3 70B and Llama 3.1 8B (August 16) on its free and developer tiers.

Llama 3.x Family (Meta) - Previous Generation

ModelParametersContextLicenseNotes
Llama 3.3 70B70B128KLlama 3.3Still widely deployed; strong general model
Llama 3.1 405B405B128KLlama 3.1Largest dense Meta model; superseded by Llama 4

Note: Llama 3.x remains in production, but Llama 4 Scout/Maverick offer better performance with lower active parameter counts thanks to MoE.

DeepSeek Family

ModelParametersContextStatusNotes
DeepSeek V4.1-Flash552B backbone (763B checkpoint); 8B active for prefill, 16B for decode (MoE)1M (384K max output)GA, September 10, 2026MIT; API name deepseek-flash. New causal encoder-decoder: the decoder's global KV is projected from encoder states, plus Compressed Sparse Attention 2 and FP4 KV, for 890 bytes of global KV per token (about 1/4 of V4-Flash, per DeepSeek). Native image input; reasoning effort 1-100. API: peak $0.30 / $1.20 per 1M, cache hit $0.006; off-peak half. Since the October 17 build (V4.1-Flash-1017, MIT weights), deepseek-flash serves 1017 at unchanged token prices. AA v5.0.1: 33 at max (the top released-weight score AA labels open), 29 at high, 19 non-reasoning; the max run used 174M output tokens against a 99.3M median and cost $0.60 per index task.
DeepSeek V4 Pro1.6T total / 49B active (MoE)1MGA (build V4-Pro-0813)Previewed April 24, 2026. Uses ~27% compute / 10% memory of V3.2 at 1M tokens. SWE-bench Verified 80.6%. NIST CAISI evaluation (May 2026) placed it ~8 months behind the US frontier. V4-Pro-0813 (MIT, August 13) bundles a DSpark speculative-decoding module in the checkpoint. API since August 16: peak $1.32 / $3.96 per 1M (cache hit $0.044), off-peak half. DeepSeek announced on September 10 that v4-pro traffic would route to V4.1-Flash from September 14, then reversed that on September 11 "in response to user demand"; V4.1-Pro had not launched as of November 2.
DeepSeek V4 Flash284B total / 13B active (MoE)1MSupersededReplaced by V4.1-Flash. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp now route to V4.1-Flash at its rates. It was $0.14 / $0.28 before the August 16 repricing.
DeepSeek-V3.2671B (MoE)128KPrevious genGeneral-purpose; 98% cache-hit discount ($0.28/$0.42 per 1M base at the time). Superseded by the V4 line for new builds.
DeepSeek-V3671B (MoE, 37B active)128KPrevious genGPT-4o level at a fraction of training cost; open weights.
DeepSeek-R1671B (MoE)128KReasoningMatched o1 on math/code; first open-source reasoning model.
DeepSeek-R1-Distill7B to 70B-ReasoningDistilled to smaller models; cost-efficient reasoning.

Key November 2026 context: DeepSeek is no longer the automatic cheap answer, and its prices now depend on the clock. The August 16 repricing moved V4 to peak and off-peak billing: peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays, and everything else, including weekends, is half price. V4.1-Flash then cut Flash prices on September 10, but at peak its $1.20 output only matches GPT-5.6 Luna and sits well above GPT-6 Luna's $0.50. Any cost model still using the pre-August flat $0.14 / $0.28 or $0.435 / $0.87 is off by roughly 2x to 4x. Cache-hit discounts remain steep (about 98% on V4.1-Flash), so cache-friendly batch work scheduled off-peak is where DeepSeek still wins. The V4-Pro reroute-then-reversal within a day is the operational lesson, and the October 17 swap of deepseek-flash to a new build repeats it: pin dated builds and alert when an endpoint starts serving something else. The same day DeepSeek opened a Search & Fetch API on its in-house search engine, with Search and Fetch each at $0.75 per 1,000 successful requests; residency reviews must treat those search queries as data leaving to DeepSeek. Concurrency caps limit fan-out too: 2,500 for deepseek-flash and 500 for deepseek-v4-pro.

Moonshot Kimi Family

ModelParametersContextNotes
Kimi K32.8T total / 104B active (MoE)1MJuly 16, 2026 (open weights July 27). Largest open-weight model to date. Always-on thinking, multimodal. $3 / $15 per 1M ($0.30 cached input). License: a custom Kimi K3 License, not Modified MIT. A Model-as-a-Service operator whose revenue (with affiliates) exceeds US$20M over 12 months needs a separate agreement before commercial use, and products above 100M MAU or US$20M monthly revenue must display "Kimi K3"; internal use and Moonshot's certified inference partners are exempt. AA v5.0 shows no score for it (the 57.1 quoted at launch and the later 44 were on older, non-comparable index versions). It was the first open model to top WebDev Arena at launch; on the October 27 Arena boards it is the top open model on the Text Arena (#19, 1488), while Kimi K3.1 Pro (Max) is the top open model on the WebDev Arena (#9, preliminary 1695, behind Claude Opus 5.5 at 1810). Since September 18 it runs on Amazon Bedrock, the first open-weight model there with explicit prompt caching (prompt_cache_breakpoint after at least 1,024 tokens, entries persist at least 30 minutes, cache hits do not count against input-TPM quotas), with the global cross-Region profile about 10% cheaper than the US one.
Kimi K3.1 Pro2.8T total / 104B active (MoE)1,048,576API October 23, 2026; weights November 4 (moonshotai/Kimi-K3.1-Pro, public and ungated, about 1.56 TB in 96 shards). MXFP4 weights with MXFP8 activations, trained quantization-aware. Reasoning effort low, high, or max (default max); multi-turn tool calls must return reasoning_content. $2 / $10 per 1M on Moonshot's API; kimi-k3.1-pro-highspeed $4 / $20 (2-3x the speed per Moonshot). AA v5.0.1: 46, the top open-weight score, though AA's page still labeled it proprietary on November 6. License: custom Kimi K3.1-Pro License (below).
Kimi K2.61T total / 32B active (MoE)-Released April 20, 2026. Modified MIT license. Closed to newly registered Kimi API users from October 31, 2026; Moonshot plans to retire all K2 series models by the end of December 2026 and points migrations to Kimi K3.1 Pro. Native video input; Agent Swarm scaling to 300 sub-agents and 4,000 coordinated steps. Tied GPT-5.5 on SWE-Bench Pro (58.6%); SWE-bench Verified ~80.2%.
Kimi K2.7 Code1T total / 32B active (MoE)256KJune 12, 2026. Coding-focused build on K2.6 (Modified MIT) with a MoonViT vision encoder. Reports about +21.8% over K2.6 on Moonshot's own Kimi Code Bench v2 with roughly 30% fewer thinking tokens (vendor benchmark). API about $0.95 / $4.00 per 1M (on the same K2 retirement path: no new users since October 31, retirement planned by the end of December 2026). Superseded by K3 as Moonshot's flagship.
Kimi K2-Thinking-0905--First model to hit 100% on AIME 2025 (reasoning variant).

Best for: Long-horizon agent workloads, video understanding, open-weight agent stack alternative to closed frontier.

K3 as a pretraining substitute: Cognition's SWE-2 (September 10, Devin-only, no weights) is post-trained from K3, and Fireworks' Ember-1 (September 23, research preview at K3's $3 / $15) is built on it, with about 40% fewer tokens than K3 (vendor-reported). Western product labs now buy an open base and compete on RL and post-training, which turns the base model's license into a supply-chain term: K3's MaaS clause is written for exactly this kind of hosted API. Moonshot then published Kimi K3.1 Pro's weights on November 4, 12 days after the API launch, under a custom Kimi K3.1-Pro License: an MIT-like permission grant with two conditions. A MaaS business whose licensee and affiliates earn more than US$20M in any consecutive 12 months needs a separate agreement with Moonshot before commercial use, and commercial products above 100M monthly active users or US$20M monthly revenue must display "Kimi K3.1-Pro" in the UI. The MaaS definition is the useful part for design reviews: giving third parties meaningful control over inference or fine-tuning (for example via API) counts, while end-user products with the capability embedded in specific features or harnesses, and mere relaying to models hosted by others, do not. Internal use that exposes nothing to third parties, and use through Moonshot's official products or certified inference partners, are exempt from both conditions. It is not an MIT license.

Alibaba Qwen Family

ModelParametersLicenseNotes
Qwen3.8-Flash-Next125B total / 6B active MoE, plus a 51B N-gram memory and a 4B MTP moduleQwen Community License 1.0August 26, 2026. An early preview of the Qwen4 architecture: Gated DeltaNet linear attention in three of every four layers plus Qwen Sparse Attention, and an N-gram lookup table that can live in host RAM rather than HBM. 262K native, 1M with YaRN; text, image, and video in. License: any MaaS or AI coding/office-assistant business needs a separate license from Qwen before commercial use, with no revenue floor; internal use is exempt. Hosted as qwen3.8-flash on Qwen Cloud at $0.15 / $0.47 per 1M.
Qwen3.8-Max2.4T total / 95B active MoEQwen3.8-Max License (gated)August 12, 2026; open checkpoint Qwen3.8-2.4T-A95B. 262K context native, extensible to ~1,010,000. Open weights under a bespoke license, not Apache: attribution required above 100M MAU or $20M monthly revenue, and a separate license from Qwen is required for Model-as-a-Service or AI Work Assistant (coding or office) businesses above US$50M aggregate revenue over 12 months. The open checkpoint is text-input-only; the API version is multimodal.
Qwen3.8-Max-0902Same 2.4T / 95BAPI onlySeptember 2, 2026 post-training refresh with no weights; US$2 / $6 per 1M on Alibaba Model Studio (international), the same as qwen3.8-max. AA v5.0: 34, above every open-weight model except Kimi K3.1 Pro (46 on v5.0.1, weights since November 4), which is the cost of staying on the open checkpoint. The open checkpoint stays at the August 12 build, so self-hosters fall behind the API model within weeks: "open snapshot, closed refresh".
Qwen3.8-27B27B denseApache 2.0August 14, 2026. The more permissive and more modality-complete artifact: accepts image and video input where the open Max checkpoint does not. 262K context, extensible to ~1M. Dense rather than MoE, which makes it the practical single-GPU option. AA v5.0: 20 (xhigh). Base for Ternary Bonsai 2; Groq moved its Qwen tier to it on September 14.
Qwen 3.6 Max-Preview~1T MoECommercial previewReleased ~April 20-27, 2026. 262K context. Topped six coding benchmarks per Alibaba.
Qwen 3.6-Plus--Released April 2, 2026. Enhanced coding.
Qwen 3.6-35B-A3B35B / 3B active MoEApache 2.0Released April 16, 2026. Open-weight workhorse.
Qwen2.5-Coder-32B32BApache 2.0Previous-generation open coding leader.
Qwen2.5-72B72BApache 2.0Previous-generation multilingual leader.
Qwen2.5-7B7BApache 2.0Efficient self-hosted option.

Within one family, the newest release (Flash-Next) carries the strictest license and the smaller 27B the most permissive one. Read the LICENSE file for each checkpoint, not the family name. Qwen4 has not shipped as of November 6: the newest language models on the Qwen Hugging Face organization are still Qwen3.8-Flash-Next and its FP8 build (August 24). Attendees reported Alibaba Cloud CTO Feifei Li saying "Qwen4 next week" at Apsara Amsterdam on October 29 (reported, not vendor-confirmed), but Alibaba has published no Qwen4 date, model card, or license, so do not plan around Qwen4 weights or assume Apache 2.0.

Mistral Family

ModelParametersContextNotes
Mistral Large 4About 1T total / 49B active (52B with embeddings and output layers), MoE1MAPI preview October 6, 2026; open weights October 29 (mistralai/Mistral-Large-4-1T-A52B). $0.70 / $2.10 per 1M, cached $0.07. License: Mistral Ecosystem Open-Weight License, under which any Model-as-a-Service business needs a separate commercial license, with no revenue threshold. The final release replaced the preview's None reasoning mode with Low, so it offers Low, Medium, and High, with High as the default: preview code that asked for no reasoning must move to low, and unset effort raises token use. AA: 26 at High, 23 at Medium (October 29). Microsoft Foundry is the first cloud to offer it, per Mistral, which also reports Terminal-Bench 4.0 41% (28% in preview) and 48% on AA's Coding Agent Index (vendor-reported). Named replacement for Medium 3.5 and Large 3.
Mistral Medium 3.5128B dense256KReleased April 29, 2026. Merges Magistral (reasoning) + Pixtral (vision) + Devstral 2 (coding) into one model. 77.6% on SWE-Bench Verified. $1.50 / $7.50 per 1M. License: Modified MIT that withdraws all rights from organizations with more than US$20M global monthly revenue (they need a commercial license or Mistral's hosted API), including derivatives. Deprecated October 31, 2026; retires April 30, 2027 (to Large 4).
Voxtral TTS4B open-weightsstreamingMarch 23, 2026 release, CC BY-NC 4.0 (non-commercial). 70ms latency, 9 languages, 3-second voice cloning. On the Mistral API, Voxtral TTS 26.03 was deprecated October 31, 2026 and retires April 30, 2027, replaced by Voxtral TTS 3 (public preview since October 20); the deprecation applies to the API model ID.
Mistral Large 3675B (MoE, 41B active)256KSparse MoE; parity with best open-weight models at launch. $0.50 / $1.50 per 1M on the API. Deprecated October 31, 2026; retires April 30, 2027. Moving to Large 4 raises the price 40%.
Mistral Small 4-256KHybrid instruct/reasoning/coding; released March 2026. $0.15 / $0.60 per 1M.
Mistral 3 (14B/8B/3B)3B to 14B-Unified family: multilingual, multimodal, Apache 2.0.
Mixtral 8x22B141B (MoE)-Previous gen; still viable for throughput.

Large 4 is Mistral's first open-weight flagship since Large 3, and its license is the strictest in this chapter for anyone selling inference: Medium 3.5's cap bites only above US$20M monthly revenue, while Large 4 needs a separate deal for any MaaS business. La Plateforme also serves Z.ai's GLM 5.3 (GA September 28; GLM 5.2 was listed to retire there October 31), so Europe's leading "sovereign" provider resells a Chinese open-weight model. Provenance and hosting jurisdiction are separate questions.

Google Gemma Family

ModelParametersContextLicenseNotes
Gemma 4 (31B dense)31B256KApache 2.0Released April 2, 2026. 140+ languages; native vision/audio; function calling.
Gemma 4 (26B-A4B MoE)26B / 4B active256KApache 2.0Sparse MoE variant.
Gemma 4 12B (Unified)11.95B dense256KApache 2.0June 3, 2026. Encoder-free multimodal: raw image patches and audio waveforms are projected into the LLM embedding space through lightweight linear layers; text, image, audio, and video (as frames) in. Official QAT int4 builds. The obvious laptop-class multimodal default.
Gemma 4 E4B8B256KApache 2.0Edge-suitable.
Gemma 4 E2B5.1B / 2.3B active256KApache 2.0Smallest variant; mobile/embedded.
DiffusionGemma (26B-A4B MoE)26B / ~4B active256KApache 2.0June 10, 2026. Google DeepMind's first open-weight text-diffusion model; denoises blocks of tokens in parallel for roughly 4x faster generation (1000+ tokens/sec on one H100). Lower quality than standard Gemma 4; aimed at low-latency and in-line editing.

Google has released no new Gemma LLM since DiffusionGemma, and Gemma 5 has not been announced.

Zhipu / Z.ai GLM Family

ModelParametersContextLicenseNotes
GLM-5.3744B total / 40B active (MoE), same base as GLM-5.21MGLM-5.3 License (MIT plus a security-review clause)Coding Plan August 14, 2026; general API August 18; open weights August 28, the date Z.ai had targeted. The entire gain comes from extended post-training rather than a new pretraining run. The license is MIT except that a Model-as-a-Service operator with more than US$10B aggregate revenue over 12 months must pass a Z.ai security review before commercial use, so almost everyone gets effectively MIT terms. API $1.40 / $0.26 cached / $4.40 per 1M. Amazon Bedrock made it GA on October 5 (1M context, up to 128K output, US and Global profiles), and AWS states inference data is not shared with the model provider, so teams barred from calling Chinese vendors directly can run it inside the AWS boundary. AA v5.0.1: 32 at max, 23 at low. NIST CAISI (September 17) called it the most cyber-capable open-weight model released to date, about four months behind the US frontier.
GLM-5.3-Flash320B total / 18B active (MoE)1M (128K max output)MITAugust 26, 2026, with none of GLM-5.3's review clause. First natively multimodal GLM-5 model, built for GUI observation and interaction. Hybrid linear and sparse attention (Z.ai: 3.01x less attention compute and a 4.44x smaller KV cache than GLM-5.3). $0.15 / $0.50 per 1M, about 1/9 of GLM-5.3. AA v5.0: 28.
GLM-5.2744B total / 40B active (MoE)1MMITCoding-plan access June 13, 2026; open weights June 16-17. Built for long-horizon agentic coding and tool use. Reports SWE-Bench Pro 62.1 (ahead of GPT-5.5 at 58.6 on that benchmark); vendor-reported. API roughly $1.40 / $4.40 per 1M. Superseded by GLM-5.3.

Best for: Open-weight agentic coding and long-horizon tool use where a 1M context and a permissive license matter. GLM-5.3-Flash is the new default self-host candidate in the 18B-active class. Verify benchmark claims on independent leaderboards.

A safety window is not a safety guarantee. Within days of the GLM-5.3 weight drop, Hugging Face carried community "uncensored" and abliterated derivatives, including a GLM-5.3-Flash build with roughly 220K monthly downloads. A staged release delays open access; it does not stop safety training from being stripped afterward.

Thinking Machines Inkling - July 2026

ModelParametersContextLicenseNotes
Inkling975B total / 41B active (MoE)1M (64K or 256K via the lab's Tinker API)Apache 2.0Released July 15, 2026: Thinking Machines Lab's first public model, pretrained on 45T tokens of text, image, audio, and video. SWE-Bench Verified 77.6%; the leading US open-weights model per Artificial Analysis at launch, with safety scores the lab reports as aligning with frontier models. NVFP4 checkpoint optimized for NVIDIA Blackwell. AA v5.0: 12.
Inkling-Small~266B total / 12B active-Apache 2.0Full weights released July 30, 2026 (post-trained beyond the earlier preview checkpoint), plus an NVFP4 build. Text, image, and audio in. Vendor-reported SWE-bench Verified 80.2%, beating its larger sibling. AA v5.0: 13.

Why it matters: A brand-new US lab shipping the leading American open-weights family, positioned explicitly as a fine-tuning foundation. Kimi K3 and Inkling landing in the same July week made it the strongest open-weights month on record at the time. By November both Inkling models trail Meta's Muse Spark Open (25), Reflection's Beam 1.0, and MBZUAI's K2 Horizon 375B (both 18) on AA v5.0.1 and sit far behind the Chinese leaders (Kimi K3.1 Pro at 46 by weight availability; DeepSeek V4.1-Flash-1017 at 33 and GLM-5.3 at 32 under AA's own label), and no new Thinking Machines model has shipped since July.

Meta Muse Spark (Closed Weights) - April 2026 STRATEGIC SHIFT

AttributeValue
LicenseClosed weights - first proprietary model from Meta Superintelligence Labs
CapabilitiesMultimodal reasoning with Instant / Thinking / Contemplating modes
ReleasedApril 8, 2026

Strategic significance: Meta's first non-open model since the original Llama era. Signals that frontier-quality work may require a closed-development feedback loop. Meta paused the Llama 4 Behemoth release at the same time amid capability concerns, and Behemoth has never shipped. The open-vs-closed equilibrium became two-tier: frontier closed models lead, and open weights catch up via distillation, RL, and ecosystem iteration. The lead was commonly put at 6 to 12 months in April; by September NIST CAISI estimated GLM-5.3 at about four months behind the US frontier on cyber, while AA v5.0.1 shows 26 points between the best callable closed model and the best model AA labels open (59 vs 33), or 13 if you count Kimi K3.1 Pro (46), whose weights shipped November 4.

July 2026 update: Muse Spark 1.1 shipped July 9 alongside the public preview of the Meta Model API, Meta's first self-serve paid API: OpenAI-compatible, $1.25 / $4.25 per 1M, roughly a quarter of rival flagship rates. Vendor-reported benchmarks lead on scaled tool use (MCP Atlas 88.1) and professional tool use (JobBench 54.7). Meta charging for API access completes the pivot away from open-weight Llama; there is no Llama 5, and Behemoth remains shelved.

August 2026 update: Muse Glimmer (August 10) is Meta's first open-weight release since Llama 4 and, at Apache 2.0, its most permissive license ever for an open model. It is a 30B dense multimodal model aimed at always-on local agent work (local coding agents, function calling, LLM-as-a-judge), with a 131,072-token context; quantized to 4-bit it fits under 20GB and runs on a single 24GB consumer GPU. Muse Spark 1.2 and Muse Code, a terminal coding agent, shipped alongside it, notable for a muse-spark-1.2-contributor tier priced at $0.10 / $0.20 per 1M (against $1.25 / $4.25 standard) in exchange for permission to train on your prompts and completions. Data-for-discount as an explicit, published API tier is new, and worth a policy decision before anyone enables it.

September 2026 update: Muse Spark 1.3 (muse-spark-1.3, September 2) keeps the $1.25 / $4.25 price ($0.15 cached) and adds a 1M-token context with text, image, video, and PDF input; the contributor tier (about $0.10 / $0.20, traffic may be used by Meta) continues. The max tier launched in limited preview pending safety testing and is now listed on Standard. AA v5.0 shows no score for it yet. It is more verbose than 1.2: The Decoder put cost per task at $0.55 versus $0.40 despite flat per-token prices, a clean example of why cost per task, not cost per token, should drive selection. Meta also launched Muse, a consumer personal agent that can browse and fill out forms (September 8).

October 2026 update: Muse Spark Open (October 15, meta-models/Muse-Spark-Open-816B-A42B) is Meta's largest open-weight release: 816B total / 42B active, 1M context, text, image, video, and audio in. It ships under the Muse Model License, not Apache: products must display "Built with Muse Spark Open", and a Model-as-a-Service operator whose revenue with affiliates exceeds US$20M over 12 months must sign a separate agreement with Meta. AA v5.0.1 scores it 25, the top US open-weight model and 8 points behind DeepSeek V4.1 Flash 1017. Meta's frontier models stay closed, so the open line still trails the closed one by design. Two operational details matter. The Hugging Face repo requires manual approval and collects legal name, date of birth, country, affiliation, job title, and IP-based location, so CI and fresh clusters cannot pull it without a human-approved account; mirror it. And it is also hosted: Fireworks serves it serverless at $1.25 / $0.15 cached / $4.25 per 1M on US-only capacity (blog post October 20), Microsoft Foundry has offered it on managed compute since October 15, and GroqCloud lists it as a preview model. That matches the closed Muse Spark 1.3's price, so the choice between them is license and data terms, not cost.

Three days earlier Meta shipped that closed frontier line: Muse Wonder (October 12, muse-wonder-1.0) at $4 input / $0.20 cached / $20 output per 1M on the Meta Model API and in Muse Code, with a 1M-token context and computer use that Meta says can operate browsers, desktop, and mobile apps. Muse Wonder Full (muse-wonder-full-1.0) has no API price at all: it is available only through Muse Code subscriptions ($50 and $200 monthly plans). Meta claims state-of-the-art results on personal and professional agentic tasks but published no benchmark numbers at launch. Artificial Analysis v5.0 scores Muse Wonder 43 and Muse Wonder Full 50, both at max effort. Two design consequences. At Opus 5.5's list price and 15 AA points below it, Muse Wonder competes on Meta's stack and computer use, not capability per dollar. And the strongest Meta model is a subscription product, so the model your API evals measure is not the best one Meta ships.

Xiaomi MiMo-V2.6 - September 2026

ModelParametersContextLicenseNotes
MiMo-V2.6-Pro1.02T total / 42B active (MoE)1MMITSeptember 21, 2026, ungated weights. Omni-modal input (text, image, video, audio). 60 sliding-window plus 10 global attention layers and a 5-layer MTP drafter that predicts 7 tokens per pass. AA v5.0: 30, just behind DeepSeek V4.1-Flash-1017 and GLM-5.3 (32). Vendor-reported Terminal-Bench 2.1 89.9, DeepSWE v1.1 71.9. OpenRouter $0.435 / $0.87 per 1M. Xiaomi's SGLang recipe serves it across two nodes with 16-way tensor parallelism.
MiMo-V2.6-Flash309B total / 15B active (MoE)1MMITSame date and modalities; fits one 8-GPU node. OpenRouter $0.14 / $0.28 per 1M.

Why it matters: a phone maker took the top open-weight spot in September and on AA v5.0.1 sits at 30, 3 points behind DeepSeek V4.1-Flash-1017 (33) and 16 behind Kimi K3.1 Pro (46, weights since November 4), and its OpenRouter prices are exactly DeepSeek's old V4 Pro and V4 Flash list prices, so it took over the price point DeepSeek vacated in August.

Other Open-Weight Families to Know (November 2026)

ModelReleasedSizeLicenseNotes
Tencent Hy4 previewAug 28, 2026770B / 49B active MoEApache 2.0 (standard text; no use-policy appendix, revenue gate, or territorial carve-out)1M context; Gated DeepSeek Sparse Attention. Vendor-reported SWE-Bench Pro 65.7. TokenHub $0.834 / $2.501 per 1M. Labeled a preview with known issues such as over-verification. Notable from a lab whose April Hy3 preview license excluded the EU, UK, and South Korea. The full Hy4 (November 4) adds image input and computer use but is API-only for now (Tencent TokenHub, OpenRouter tencent/hy4); Tencent says weights and a system card will follow, with no date or license, and AA v5.0.1 scores it 35 and labels it proprietary. The Apache 2.0 terms cover the preview checkpoint only
MBZUAI K2 HorizonSep 3, 2026Six models, 0.9B to 375B-A23BApache 2.0, with training data, code, intermediate checkpoints, and logs512K context (0.9B: 128K). The 375B-A23B scores 18 on AA v5.0, above Inkling and Nemotron 3 Ultra. The best fully reproducible family for provenance and contamination audits
IBM Granite 4.2Aug 25, 20263B, 8B, 30B denseApache 2.0128K native, extendable to 512K; full, non-thinking, and low-effort modes; official FP8, MXFP4, NVFP4, GGUF, and MLX builds. The common answer to "US-origin, Apache, one GPU"
IBM Granite Switch 4.1Oct 14, 2026 (rev. 4.1.1 Oct 27)8B, 30BApache 2.0One endpoint packaging a Granite 4.1 base, two aLoRA adapters (guardian-core for safety detection, decision for scoring labeled options), and a 470M Granite Speech 5.0 encoder; 131,072-token context. Its optional vLLM backend is pinned to ">= 0.26.0, < 0.31.0", below this guide's vLLM 0.32.0 security floor, a concrete reason to check model-card engine pins in CI
NVIDIA Nemotron 3 UltraJun 4, 2026550B / 55B activeOpenMDW-1.1Not 500B. AA v5.0: 12. Superseded as the NVIDIA-stack open default by Nemotron 3.5 Super VL (below). A Labs competitive-coding variant (Sep 3) reached IOI 2026 gold-level scores in an unofficial run using traces distilled from Z.ai's GLM-5.2, so distillation lineage now crosses borders both ways. Nemotron 4 is reportedly still in training; NVIDIA has given no date
MiniMax-M3 / M3.1-FlashJun 1 / Oct 23, 2026~428B / 23B active (M3.1-Flash per its model card)MiniMax Community License (gated; see the licensing table)1M context with MiniMax Sparse Attention. MiniMax-M3.1-Flash weights landed October 23 (AA v5.0: 29); the full M3.1 is still closed
NVIDIA Nemotron 3.5 Super VLOct 15, 2026120B / 12B active, Mamba-2 and Transformer hybrid MoE with MTPOpenMDW-1.1Text, image, and up to 2-minute video in, up to 1M context; BF16 and NVFP4 checkpoints; the vLLM v0.32.0 recipe serves it on 4x H100 or 2x GB200. AA 14 vs 12 for Nemotron 3 Ultra; NVIDIA-reported NeMo Gym results include SWE-bench Verified 71.10 and Terminal-Bench 2.1 60.67. The new NVIDIA-stack open default: higher score, multimodal, four H100s
Reflection AI Beam 1.0Weights Oct 16, 2026501.3B / 23.4B active, 256 expertsApache 2.01M context, text only; FP8, NVFP4, and a DFlash2 speculative drafter. AA v5.0.1: 18, tied with K2 Horizon 375B and second among US open-weight models behind Muse Spark Open. Reflection reports SWE-bench Verified 81.6% and Terminal-Bench 2.1 80.1% (BF16, temperature 0.7). A second unconditional Apache 2.0 US option at 500B scale: a license and provenance pick, not a capability pick
Ant Group Ling-3.1-flash-1019Oct 2026 (scored Oct 22)560B / ~25B activeMIT1M context; hybrid linear MoE with MTP. AA v5.0.1: 30, tied with MiMo-V2.6-Pro, but it used 240M output tokens on the index against a 99.3M median, so cost per task, not AA's $0.30 / $0.90 list, decides it
Nanbeige4.3-33B (BOSS Zhipin)Oct 24, 202633B denseApache 2.0Context up to 524,288; FP8, INT8, and INT4 builds October 31. AA v5.0.1: 20, tying Qwen3.8-27B at xhigh, but at 160M output tokens against a 77.7M median; no public API. A single-GPU Apache peer to Qwen3.8-27B at higher cost per task
Ricursive Ripple-65B PreviewNov 4, 202665B denseApache 2.0Up to 1M context; pretrained on 12T tokens; ungated. Vendor-reported Terminal-Bench 2.1 80.52, topping the vendor's own comparison with Beam (80.1) and Muse Spark Open (76.4); AA has not scored it. Shortlist only after an independent score appears
Ant Group Ling-3.0-flash-VL / LLaDA2.2-miniSep 4 / Sep 5, 2026124B / 5.5B active; 16B / 1.4B activeMIT; Apache 2.0A GUI-agent vision-language MoE, and a diffusion LM with DELETE and INSERT edit tokens and a 128K context
StepFun Step 5 PreviewAPI Sep 20; weights around Oct 15, 2026600B / 27B activeStepFun Model License 1.0API $1 / $2.70 per 1M; AA v5.0: 29. Verbose: about 160M output tokens on AA's earlier suite vs an 81M median, which eats much of its per-token advantage. The license gates MaaS above US$20M revenue and any branded device (phones, cars, robots), and counts models trained substantially on its outputs as derivatives

The Open-Weight Licensing Split (August to November 2026)

August was the month open weights stopped meaning one thing, September made the pattern clear, and October tightened it: gated licenses converge on one trigger, whether you sell inference or an AI assistant, and Mistral Large 4 dropped the revenue floor entirely. The license is as much a design input as the benchmark:

ModelReleasedSizeLicense postureWhat the license actually does
Mistral Large 4Oct 29 (weights)~1T / 49B activeMistral Ecosystem Open-Weight LicenseAny Model-as-a-Service business needs a separate commercial license, with no revenue threshold. The strictest MaaS gate on a flagship-class open model, and it applies to the named replacement for Medium 3.5 and Large 3
Kimi K3.1 Pro (Moonshot)Nov 4 (weights)2.8T / 104B activeKimi K3.1-Pro LicenseMIT-like grant; separate agreement for MaaS businesses above US$20M revenue over 12 months, where MaaS means giving third parties meaningful control over inference or fine-tuning; "Kimi K3.1-Pro" UI credit above 100M MAU or US$20M monthly revenue; internal use and certified partners exempt
Muse Spark Open (Meta)Oct 15816B / 42B activeMuse Model License"Built with Muse Spark Open" attribution in any product; a separate agreement with Meta for MaaS operators above US$20M revenue (with affiliates) over 12 months; internal use and building it into your own product's features are excluded. The Hugging Face repo is gated by manual approval and is not intended for individuals under 18
Step 5 Preview (StepFun)Weights around Oct 15600B / 27B activeStepFun Model License 1.0Separate agreement for MaaS businesses above US$20M revenue in any 12 months, and for any branded hardware that exposes the model (phones, vehicles, wearables, robots) with no threshold; attribution above 100M MAU or US$20M monthly revenue; models trained substantially on its outputs count as derivatives
MiniMax-M3.1-FlashOct 23~428B / 23B activeMiniMax Community LicenseRights for non-commercial purposes; commercial use (fee-based products relying on the model, API offerings including cloud-hosted ones, and commercial deployment of fine-tuned derivatives) only under conditions: prior written authorization above US$20M yearly revenue, a one-time notice to MiniMax for everyone else, and "Built with MiniMax M3.1-Flash" display for all commercial users. An appendix bars uses including any military purpose, and there is no explicit internal-use exemption. The full M3.1 stays closed
Qwen3.8-Flash-Next (Alibaba)Aug 26125B / 6B activeQwen Community License 1.0Any MaaS or AI coding/office-assistant business needs a separate license before commercial use, with no revenue floor; internal use exempt; attribution above 100M MAU or US$20M monthly revenue. Stricter than the flagship's license
Qwen3.8-Max (Alibaba)Aug 122.4T / 95B activeBespoke, commercially gatedFree use, modification, and resale, but attribution is required above 100M MAU or $20M monthly revenue, and a separate license from Qwen is required to run a Model-as-a-Service or AI Work Assistant (coding or office productivity) business when aggregate revenue with affiliates exceeds US$50M over any 12 consecutive months
Kimi K3 (Moonshot)Jul 27 (weights)2.8T / 104B activeKimi K3 LicenseSeparate agreement for MaaS operators above US$20M revenue over 12 months; attribution above 100M MAU or US$20M monthly revenue; internal use and certified partners exempt
Mistral Medium 3.5Apr 29128B denseModified MITA blanket revenue cap, not a MaaS trigger: no rights at all above US$20M global monthly revenue. Deprecated on Mistral's API October 31 (retires April 30, 2027), though the weights stay under this license
GLM-5.3 (Z.ai)Aug 28 (weights)744B / 40B activeGLM-5.3 License (MIT plus review)Security review before commercial use only for MaaS operators above US$10B aggregate revenue; effectively MIT for almost everyone. Weights shipped on the stated date after a two-week safety window
Qwen3.8-27B (Alibaba)Aug 1427B denseApache 2.0No conditions. The smaller sibling is both more permissive and more modality-complete than the flagship
Muse Glimmer (Meta)Aug 1030B denseApache 2.0No conditions; Meta's most permissive open license to date
Tencent Hy3Global Aug 5 (model Jul 6)295B / 21B activeApache 2.0Fully permissive with no geographic carve-outs, reversing the April preview's restrictive license that excluded the EU, UK, and South Korea
Tencent Hy4 previewAug 28770B / 49B activeApache 2.0No conditions; verbatim Apache text with no use-policy appendix. Applies to the preview checkpoint only: the full Hy4 (November 4) has no weights or license yet
MiMo-V2.6-Pro / Flash (Xiaomi)Sep 211.02T / 42B; 309B / 15BMITNo conditions
DeepSeek V4.1-Flash / V4.1-Flash-1017Sep 10 / Oct 17552B backboneMITNo conditions; the 1017 build now serves deepseek-flash
Ling-3.1-flash / -1019 (Ant Group)Oct 2026560B / ~25B activeMITNo conditions
Beam 1.0 (Reflection AI)Oct 16 (weights)501B / 23B activeApache 2.0No conditions
Nanbeige4.3-33B, Ripple-65B PreviewOct 24 / Nov 433B dense; 65B denseApache 2.0No conditions
GLM-5.3-Flash (Z.ai)Aug 26320B / 18B activeMITNo conditions; none of GLM-5.3's review clause
K2 Horizon (MBZUAI)Sep 30.9B to 375BApache 2.0No conditions, plus open training data and code
Ling-3.0-flash / tiny (Ant Group)Aug 5 / Aug 11124B / 5.1B active; 7.9B / 1.3B activeMITNo conditions
Nemotron 3.5 Lightning / Super VL (NVIDIA)Aug 11 / Oct 1530B / 3B active; 120B / 12B activeOpenMDW-1.1Linux Foundation license, free for commercial use
Qwen3.8-Max-0902, Tencent Hy4 (full)Sep 2 / Nov 4-Weights not releasedAPI-only; do not plan around weights or a license until they ship

Three things follow for anyone building on open weights. First, read the license before the model card: the trigger is almost always "are you selling inference or an AI assistant", which is exactly the business many teams (and interview candidates) propose building, and a gated license can make an "open" flagship unusable for it while its smaller sibling is unencumbered. Second, staged release is now a working precedent, and so is the closed refresh: Z.ai shipped GLM-5.3 commercially, held the weights for a two-week safety window, and released them on the date it named; StepFun's Step 5 Preview weights followed its API by about four weeks; Moonshot's Kimi K3.1 Pro weights followed its API by 12 days (November 4); Alibaba's Qwen3.8-Max-0902 refresh and the full Tencent Hy4 remain API-only, so do not build a roadmap on weights that have not shipped. Third, mirror what you depend on: NVIDIA agreed to buy Hugging Face for $12.93B (confirmed September 3; no closing date disclosed), so once the deal closes, the main open-weight distribution hub will be owned by the company selling the GPUs those pipelines run on. Keep internal mirrors of the exact weights and license files you deploy; gated repos such as Muse Spark Open, which needs a human-approved Hugging Face account, make that mandatory for CI and fresh clusters.

Small and On-Device Models (November 2026)

ModelSizeContextLicenseNotes
Nemotron 3.5 Bolt 74B-A9B 24GB (NVIDIA)74B / 9B activeUp to 256KOpenMDW-1.1October 15. Compressed from Nemotron 3.5 Super VL with Star Elastic into a 24GB packed checkpoint: experts in IQ2-XXS (about 2.06 bits per element), Mamba layers in NVFP4, embedding and output layers in Q8_0, with accuracy recovered by quantization-aware distillation. Runs on llama.cpp, Ollama, LM Studio, and vLLM; targets RTX PCs, RTX Spark laptops, and DGX Spark. A companion Bolt E4B 4GB followed October 23
Ling-3.0-tiny (Ant Group)7.9B / 1.3B active256KMITGenuinely laptop-class agentic model: roughly 8.3 GiB peak memory at 8K context, 86-90 tok/s on an M4 Pro MacBook at FP8
LFM2.5-2.6B (Liquid AI)2.6B128KOpen weightsBuilt for on-device tool calling and multi-step planning rather than chat; under 2.5 GB memory, runs down to a Raspberry Pi
LFM2.5-VL-3B (Liquid AI)3B-Open weightsOn-device vision-language model for screen understanding and GUI grounding (August 12)
Shieldstral 1.0 (Mistral)3B32KApache 2.0Policy-adaptive multimodal safety classifier (text and image, 12 languages): moderation policies are supplied in natural language at inference time, so a policy change needs no retraining. Runs on one 16GB GPU
Muse Glimmer (Meta)30B dense131KApache 2.0Under 20GB at 4-bit; targets always-on local agents
Gemma 4 12B Unified (Google)11.95B dense256KApache 2.0Encoder-free multimodal (text, image, audio, video frames); official QAT int4 builds
Granite 4.2 3B / 8B (IBM)3B, 8B dense128K (512K extended)Apache 2.0Official FP8, MXFP4, NVFP4, GGUF, and MLX builds
K2 Horizon 0.9B / 3.7B / 7B (MBZUAI)0.9B to 7B128K (0.9B), 512KApache 2.0MBZUAI says the 0.9B is small enough for a watch; GGUF builds at launch
Ternary Bonsai 2 27B (PrismML)27.36B at 1.72 bits/weight262KApache 2.0Derived from Qwen3.8-27B. A 5.95 GB file vs ~54 GB at FP16; PrismML reports 98.2% of FP16 quality across 14 thinking-mode benchmarks. Needs PrismML's llama.cpp and MLX forks (stock llama.cpp cannot load it)
Needle 3 (Cactus Compute)121M-Apache 2.0On-device tool calling, structured extraction, and embeddings in an 8-29 MB file at about 2 bits per weight
Apple Foundation Models 3 (on-device)3B dense; 20B sparse activating 1-4B-OS-providedThe 20B Core Advanced model keeps its full weights in NAND flash and loads selected experts into DRAM; ships in the iOS 27 cycle

The shift: the small-model tier stopped competing on chat quality and started competing on tool calling, planning, and screen grounding, which are the capabilities a local agent actually needs. September added two hardware-relevant patterns: ternary-native weights that put a 27B reasoner in about 6 GB, and flash-resident MoE that keeps most of a model out of DRAM. October added a third: NVIDIA's Bolt packs a 74B-total MoE into a 24GB checkpoint, so "a 70B-class model needs about 40GB" no longer holds. All three change the sizing math in On-Device and Edge Deployment.


Specialized Models

Coding (November 2026)

SWE-bench Verified and the public SWE-Bench Pro split are saturated (Opus 5 scores 99.4% on the Pro v2 public split versus 81.6% on the private set), so cite Terminal-Bench 4.0 and SWE-Bench Pro v2 private, with effort and runner.

ModelRoleEvidence
Claude Opus 5.5Default agentic coding pick at $4/$20First on the Terminal-Bench 4.0 leaderboard at 64.85% (max; tbench.ai, entries added October 6); 66.4% at xhigh in Anthropic's run (vendor-reported); Claude Code's default Opus since 2.1.280
Claude Sonnet 5.5Workhorse coding at $2/$10Second on the Terminal-Bench 4.0 leaderboard at 61.82% (max); 70.6% in Anthropic's run (vendor-reported); Claude Code's Sonnet default since 2.1.284
GPT-6 AstraOpenAI's ceilingTerminal-Bench 4.0 58.18% at max (leaderboard), tied with GPT-6.1 Sol; about a third of GPT-5.6 Sol's tokens per coding task (AA)
Claude Fable 5.1Ceiling on the Claude sideTerminal-Bench 4.0 57.88% at max (leaderboard); 55.8% in Anthropic's run
GPT-6.1 SolOpenAI's $2/$10 coding defaultTerminal-Bench 4.0 58.18% (leaderboard), tied with Astra at a fifth of its list price; Codex CLI default since rust-v0.159.1 (September 29); AA v5.0 55 (max)
GPT-6.1 TerraOpenAI's $1/$5 middle tierAA v5.0 51 at max but 38 at the default medium; test at the effort you will pay for
Gemini 3.8 FlashCheapest near-frontier coderDeepSWE v1.1 tie near 74% at $2.36 per task (vs Astra xhigh $4.43, Opus 5 max $11.84); SWE-Bench Pro v2 private 211/272
Claude Opus 5Prior Claude flagshipSWE-Bench Pro v2 private 222/272 (81.6%), the top private score Scale published; Terminal-Bench 4.0 53.94% at xhigh
Grok 4.7Cursor-native optionTerminal-Bench 4.0 37.58% at xhigh (leaderboard); on all Cursor plans
Kimi K3Largest open modelSWE-Bench Pro v2 private 214/272; custom license gates MaaS use
Claude Haiku 5.5Cheap coding sub-agentTerminal-Bench 4.0 39.2% (Anthropic-reported) at $0.10/$0.50; keep each request under 100K tokens or it bills 5x
GLM-5.3 / GLM-5.3-FlashOpen agentic codingGLM-5.3 SWE-Bench Pro v2 private 211/272; Flash is MIT at $0.15/$0.50
Kimi K3.1 Pro / DeepSeek V4.1-Flash-1017 / MiMo-V2.6-ProTop open models overallAA v5.0.1 46, 33, and 30; K3.1 Pro weights since November 4 under a MaaS-gated custom license, the other two MIT

Reasoning & Math

ModelApproachBest For
GPT-6 AstraEffort low to max, no noneARC-AGI-3 62.7% on ARC Prize's Standard harness (max, $26,098) and 99.9% on the Provider Adapter harness (high, $18,817); HLE-Diamond without tools 59.9% vs Opus 5.5 54.6%
Claude Opus 5.5 / Fable 5.1Always-on adaptive thinkingAnthropic's HLE with tools: Opus 5.5 67.7% vs Astra 57.2% (vendor-reported). On HLE-Diamond with tools the order flips (Astra 82.9%, Opus 5.5 73.9%), so pick the set that matches your task
Gemini 4 ArgonThinking model (output price includes thinking)AA v5.0 50 (high); budget at $4/$20 from December 1
Gemini 3.1 Pro Deep ThinkSustained chain-of-thoughtScientific reasoning on the preview Pro model
DeepSeek V4.1-FlashEffort dial 1-100Open (MIT) reasoning at peak $0.30/$1.20; AA v5.0.1 33 at max but 29 at high, so price the effort you serve
DeepSeek-R1RL-based thinkingHistorical open reasoning baseline
Grok 4.3 (DeepSearch)Web-grounded reasoningResearch tasks needing live information

GPQA Diamond is saturated (Astra about 96%), so it no longer separates frontier models.

Long Context (1M+)

ModelWindowPricing past the threshold
Llama 4 Scout10MSelf-hosted; industry-leading open-weight window
Claude Opus 5.5 / Sonnet 5.5 / Fable 5.11MFlat to 1M on Claude 4.6 and later, including fast mode
Claude Haiku 5.51MThe exception: a prompt over 100K tokens bills the whole request at 5x ($0.50 / $2.50), and cached tokens count toward the threshold
GPT-6 Astra / Sol / Luna, GPT-6.1 Sol / Terra1.05MThe whole request bills at long-context rates (2x input, 1.5x output) above 272K input
Gemini 4 Argon1,048,576 shared in and outNo prompt-length tiers; output defaults to 65,536 and reaches about 1M only through long decode
Gemini 3.8 Flash / 3.1 Pro Preview~1M input3.1 Pro Preview moves to $4/$18 above 200K
Grok 4.7500KEvery token doubled once the prompt reaches 200K
MiMo-V2.6, GLM-5.3, DeepSeek V4.1-Flash, Hy4 preview1MOpen weights; KV cost per token varies by orders of magnitude across architectures

Embedding Models

API Embedding Models (November 2026)

ModelMax inputCost/1MNotes
OpenAI text-embedding-3-large8,192$0.13Unchanged; OpenAI shipped no new embedding model in 2026, and there is no "text-embedding-4"
OpenAI text-embedding-3-small8,192$0.02Cheapest mainstream API option
Voyage voyage-4-large / voyage-4 / voyage-4-lite32K$0.12 / $0.06 / $0.02January 15, 2026; one shared embedding space across the three sizes
Voyage voyage-context-4, voyage-code-4-$0.12voyage-code-4 +27.54% over voyage-code-3 (vendor-reported)
Cohere embed-v5.0-pro / embed-v5.0-fast128K$0.12 / $0.08September 30, 2026; shared embedding space
Google gemini-embedding-2-See Google's pricing pageStable April 2026; multimodal. Replaces text-embedding-004 (shut down January 14, 2026)

The MTEB scores this table used to carry were 2024-25 numbers that do not compare with current models. Prefer RTEB (retrieval-focused, with private sets) plus an eval on your own corpus.

Open Source Embedding Models

ModelDimensionsMax TokensMTEB (legacy)Notes
Nemotron 3 Embed 8B / 1B (NVIDIA)---July 16, 2026, OpenMDW-1.1; RTEB Multilingual #1 at 78.5 (vendor-reported)
pplx-embed-v2-context-9b-preview (Perplexity)---September 25, 2026 preview; do not mix its vectors with later releases
EmbeddingGemma 2 (Google)---October 2026, 740M parameters, Apache 2.0; text, image, video, and audio in one embedding space
BGE-large-en-v1.5102451263.9Instruction-tuned
E5-mistral-7b-instruct40963276866.6Strong with instructions
Nomic-embed-text-v1.5768819262.3Long context, open
GTE-Qwen2-7B358432K72.1Strong 2024 open baseline

Embedding Selection Guide

RequirementRecommendedWhy
Best qualityvoyage-4-large or embed-v5.0-pro, validated on your corpusCurrent flagship APIs; benchmark on RTEB, not MTEB alone
Cost-efficientvoyage-4-lite or text-embedding-3-small$0.02/1M
Self-hostedNemotron 3 Embed 8BOpen (OpenMDW-1.1); RTEB Multilingual leader (vendor-reported)
Long documentsembed-v5.0 (128K) or voyage-4 (32K)Long input windows
MultilingualNemotron 3 EmbedRTEB Multilingual leader (vendor-reported)
Multimodalgemini-embedding-2Multimodal input in one model

See Embeddings and Vector Spaces for how to evaluate them.


Model Selection Framework

Decision Tree

text
What is your primary constraint?
├── Cost → Use smaller model, consider open source
│ ├── Very cost sensitive → GPT-6 Luna ($0.10/$0.50), Claude Haiku 5.5 (same price under 100K-token prompts), Gemini 3.8 Flash (budget on the 2027 price), DeepSeek V4.1-Flash (off-peak)
│ └── Moderate budget → GPT-6.1 Terra ($1/$5), Claude Sonnet 5.5, GPT-6.1 Sol ($2/$10), Muse Spark 1.3
│
├── Quality + Reasoning → Use frontier models
│ ├── Highest reasoning → Claude Fable 5.1, GPT-6 Astra (xhigh/max), Claude Opus 5.5 at high effort
│ └── Coding + reasoning → Claude Opus 5.5 (start here), Claude Sonnet 5.5, GPT-6.1 Sol
│
├── Latency → Use fast models or paid speed tiers
│ ├── Interactive and cheap → GPT-6 Luna, Claude Haiku 5.5 at low effort, Gemini 3.8 Flash at low thinking, GPT-6.1 Terra at `none`
│ └── Frontier quality, faster → Opus 5.5 fast mode ($8/$40), GPT-6 Astra Fast (2x) or Ultrafast (6x)
│
├── Self-hosting → Use open models (read the license first)
│ ├── Maximum capability → Kimi K3.1 Pro (MaaS gate above $20M), DeepSeek V4.1-Flash-1017 (MIT), GLM-5.3 (MIT plus review above $10B), MiMo-V2.6-Pro (MIT); Mistral Large 4 gates any MaaS use
│ ├── Good balance → GLM-5.3-Flash (MIT), Tencent Hy4 preview (Apache 2.0, preview checkpoint only), Qwen3.8-27B (Apache 2.0)
│ └── Edge/mobile → Gemma 4 12B or E4B, Granite 4.2 3B/8B, K2 Horizon 0.9B-7B
│
└── Privacy → Self-host, ZDR, or residency-pinned endpoints
└── Opus 5.5 and Sonnet 5.5 support ZDR; Fable 5.1 needs authorization; residency costs ~10% more

Semantic Routing

Static decision trees are being replaced by Semantic Routers:

  • How it works: A small, fast embedding model embeds the query. If it matches a "known easy" cluster, route to a cheap model (GPT-6 Luna, Claude Haiku 5.5 for prompts under 100K tokens, Gemini 3.8 Flash, DeepSeek V4.1-Flash off-peak). Middle clusters go to GPT-6.1 Terra or Sonnet 5.5. If it hits an "agentic/logic" cluster, route to Claude Opus 5.5 or GPT-6.1 Sol at higher effort, and reserve Fable 5.1 or GPT-6 Astra for what those still fail.
  • Benefit: Automates cost-optimization without hardcoded rules.
  • Implementation: Tools like semantic-router (Python), vLLM Semantic Router, or custom Weaviate/Pinecone classifiers.

What September and October 2026 changed for routers:

  • Prompt length is a price input, now on every major vendor. Claude Haiku 5.5 bills the whole request at 5x once the prompt passes 100K tokens (cache reads count), OpenAI bills it at long-context rates above 272K input, xAI doubles every token once a prompt reaches 200K, and the other Claude 4.6-and-later models stay flat to 1M. Measure prompt length with the target model's tokenizer and route long prompts by price, not just by window size.
  • Speed is a tier, not a model. OpenAI Fast (2x) and Ultrafast (6x on Astra and GPT-6.1 Sol), Claude fast mode ($8/$40 on Opus 5.5, out of preview since October 15), and Gemini Priority (about 1.8x) mean a latency SLO maps to a tier per route.
  • Conversation state is model-bound on Claude. Fallback chains that replay a Claude 5.5 or Fable 5.1 conversation to another model lose the reasoning blocks, and history rewrites before a thinking block fail with a 400. Fall back at turn boundaries and keep history append-only.
  • Safety refusals are routing events. Fable 5.1 refusals come back as HTTP 200 with stop_reason: "refusal"; handle them with fallbacks: "default" or your own retry, and remember that benchmark scores labeled "with fallback" already include that routing.
  • IDs do not pin behavior. GPT-6 Sol and Luna changed behavior under the same ID on September 25; Google routed gemini-3.7-flash to 3.8 Flash and gemini-3.5-flash to 3.6 Flash on October 8; deepseek-flash moved to a new build on October 17; retired xAI slugs redirect to grok-4.3 at its rates; Codex CLI and Claude Code swapped default models in point releases; from December 1 Google routes gemini-2.5-pro to 3.8 Flash and gemini-2.5-flash to 3.5 Flash-Lite; xAI now serves grok-imagine-image-pro requests with its low-quality 2.0 tier. Verify the served model, alert on changes, and rerun evals on vendor changelog events.
  • SDK upgrades change the API contract. Anthropic's eight new SDK and CLI majors send API version 2026-10-21, which rejects sampling parameters on every model. Pin SDK majors and API versions per route, and upgrade them like a model migration.
  • Gated access is a request field. OpenAI's access_programs.cyber (standard, daybreak_blue, daybreak_red) decides refusal treatment per request, and its default changed with gpt-6.1-terra. A router that swaps models without setting it explicitly can change refusal behavior or hit a 400 or 403.

Sovereign AI and Data Residency

The 2026 Regulatory Reality: Enterprises must comply with GDPR (EU), DPDPA (India), Saudi Arabia PDPL, and sectoral rules. "Sovereign AI" is now a product category.

SolutionProviderUse Case
Azure Government/SovereignMicrosoftDedicated infra in 40+ regions; approved for US Gov/EU NIS2
AWS Sovereign CloudAmazonPhysically isolated VPCs; GDPR-safe EU regions
Google Distributed CloudGoogleAir-gapped on-prem Gemini deployment
Residency-pinned API endpointsAnthropic (inference_geo: "us", 1.1x on Claude 4.6 and later, stacking on Haiku 5.5's over-100K prices), OpenAI (regional endpoints, +10% for models released from March 5, 2026), Bedrock and Google Cloud regional Claude, Bedrock regional GPT-6 Astra, Mistral EU (each about +10%). On Google Cloud (Google managed), regional endpoints serve Claude Sonnet 4.6 and earlier only; Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 need the global endpoint or the us or eu multi-region endpoints (10% over global), so residency there is geography-level, not single-regionResidency without self-hosting, priced as a published line item
Private Llama 4 / 3.3Meta (self-host)Maximum data sovereignty; open weights (Llama 4 MoE or 3.3 dense)
Permissive open weights (self-host)Z.ai GLM-5.3-Flash, Tencent Hy4 preview, Xiaomi MiMo-V2.6, IBM Granite 4.2MIT or Apache 2.0 weights; no data leaves your infra
DeepSeek (self-host)DeepSeek (MIT weights)Open weights; no data leaves your infra
Mistral Large 3 (self-host)Mistral (Apache 2.0)675B MoE; open weights; strong multilingual. Its API retires April 30, 2027, and the Large 4 weights that replace it carry a MaaS gate

Tradeoff: Dedicated sovereign clouds are priced by contract and usually carry a premium over standard regions, while per-request residency on the major APIs is now a published ~10% premium. Price residency per tenant and route only residency-bound traffic to pinned geos. Two newer wrinkles matter in regulated designs. Mistral's EU platform now serves Z.ai's GLM 5.3, so model provenance and hosting jurisdiction are separate questions. And frontier zero data retention now comes with monitoring options: Anthropic's Enterprise Frontier Safeguards (announced September 1, rolling out in phases later this fall) let customers opt in to keeping misuse-monitoring data in their own cloud account under customer-managed keys, with fully automated review, and OpenAI began previewing Private Safety Processing with select customers on August 19.

Cost Comparison at Scale (November 2026)

Assume 1M requests/day, 1K input + 500 output tokens, Standard tier list prices, no caching:

ModelInput Cost/DayOutput Cost/DayTotal/Month
GPT-6 Astra$10,000$25,000$1,050,000
Claude Opus 5.5$4,000$10,000$420,000
Gemini 4 Argon (from December 1)$4,000$10,000$420,000
Meta Muse Wonder$4,000$10,000$420,000
Gemini 3.1 Pro Preview$2,000$6,000$240,000
Claude Sonnet 5.5$2,000$5,000$210,000
GPT-6 Sol / GPT-6.1 Sol$2,000$5,000$210,000
Gemini 3.8 Flash (from January 1, 2027)$1,500$3,750$157,500
Grok 4.7$2,000$3,000$150,000
GPT-6.1 Terra$1,000$2,500$105,000
Muse Spark 1.3$1,250$2,125$101,250
Gemini 3.8 Flash (intro, through December 31)$750$1,875$78,750
DeepSeek V4.1-Flash (all at peak)$300$600$27,000
GPT-6 Luna / Claude Haiku 5.5$100$250$10,500

How to read it:

  • Cost per task beats cost per token. Verbose models eat their price advantage (Muse Spark 1.3 at $0.55 vs $0.40 per task for 1.2; Step 5 Preview at about twice the median output tokens on AA's suite), and token-efficient ones narrow theirs (Astra uses about a third of GPT-5.6 Sol's tokens per coding task). Measure tokens per task on your workload before trusting this table.
  • Caching reorders it. With a cached system prompt, cache reads cost 0.1x on most models, 0.05x on Opus 5.5, Sonnet 5.5, GPT-6.1 Sol, and GPT-6.1 Terra, and 0.025x on Fable 5.1, so per-model break-even matters more than the headline rate. Gemini 4 Argon caches only implicitly, so its cached cost is whatever hit rate Google's cache gives you.
  • Prompt size reorders it too. At 1K-token prompts Haiku 5.5 ties GPT-6 Luna; at 150K-token prompts it bills $0.50 / $2.50 on every token while Luna stays at $0.10 / $0.50 until 272K.
  • Promotions expire. Gemini 4 Argon doubles on December 1, Gemini 3.8 Flash on January 1, 2027, and OpenAI dates GPT-5.6 Sol's $4/$20 promotion to November 30, 2026, with no stated price after it.
  • Self-hosting is priced in GPU-hours, not tokens. The Silicon Data rental index on November 5 put an H100 at $2.97 and a B200 at $6.70 per GPU-hour (B200 up about 14% in four weeks), and Lambda's on-demand prices rise about 15% on November 16 (scheduled); see Pricing and Costs for the break-even math.

Capability Comparison

Benchmark Snapshot (November 2026)

ModelAA Index (v5.0.1 where marked)Terminal-Bench 4.0SWE-Bench Pro v2 private (of 272)Notes
Gemini 4 Carbon (not available to developers)60 (high, v5.0.1); 59 on v5.0--Fairwind cyber defenders only; no model ID
Claude Opus 5.559 (max, v5.0.1); 58 on v5.0 (with fallback)64.85% max (leaderboard, first); 66.4% xhigh (Anthropic)-OSWorld 2.1 partial 81.8% (Anthropic)
GPT-6 Astra57 (max, v5.0.1)58.18% max (leaderboard, tied third)-ARC-AGI-3 62.7% Standard vs 99.9% Provider Adapter (ARC Prize)
GPT-6.1 Sol55 (max, v5.0.1)58.18% (leaderboard, tied third)-$2/$10
Claude Sonnet 5.554 (max, with fallback); 41 at high61.82% max (leaderboard, second); 70.6% (Anthropic)-GDPval-AA v2.1 within 2 Elo of Opus 5.5 (Anthropic)
GPT-6.1 Terra51 (max); 38 at medium--$1/$5
Gemini 4 Argon50 (high)--DeepSWE v1.1 77.9% (Google)
Claude Fable 5.150 (xhigh, with fallback); 49 at max57.88% max (leaderboard); 55.8% (Anthropic)-Cache reads 0.025x
Meta Muse Wonder Full50 (max)--Muse Code subscription only; no API
Kimi K3.1 Pro (open weights since November 4)46 (v5.0.1)--AA still labels it proprietary; custom MaaS-gated license
Meta Muse Wonder43 (max)--$4/$20, Opus 5.5's list price
Claude Haiku 5.536 (max)39.2% (Anthropic)-5x on the whole request above a 100K-token prompt
Grok 4.735 (high)37.58% xhigh (leaderboard)-$2/$6
DeepSeek V4.1-Flash-1017 (open)33 (max, v5.0.1); 29 at high--MIT
GLM-5.3 (open)32 (max, v5.0.1); 23 at low-211 (77.6%)MIT plus MaaS review
Gemini 3.8 Flash31 (high)-211 (77.6%)Intro price through December 31
MiMo-V2.6-Pro (open)30 (v5.0.1)--MIT; Ling-3.1-flash-1019 also 30
GPT-6 Luna28 (max)--$0.10/$0.50
Kimi K3 (open)--214 (78.7%)Custom license
Claude Opus 5-53.94% xhigh (leaderboard)222 (81.6%)99.4% on the saturated public split
Inkling (open)12-184 (67.6%)Inkling-Small scores 13

Sources: Artificial Analysis Intelligence Index v5.0 (announced October 27; Carbon added October 28) and v5.0.1 (November 3, which regraded every model on CritPt v2; unmarked AA figures are v5.0), the tbench.ai Terminal-Bench 4.0 leaderboard (entries added October 6), Scale's SWE-Bench Pro v2 (September 22), ARC Prize (September 3). Vendor figures are labeled with the vendor's name. Rules for reading any table like this:

  • Index versions do not mix. Label every AA number with its version: v5.0.1 (November 3) moved Carbon from 59 to 60, Opus 5.5 from 58 to 59, and DeepSeek V4.1 Flash 1017 from 32 to 33. AA v5.0 has four equally weighted categories (Knowledge Work, Coding, Science, General) with about 60% of its questions on private or held-out sets, and its scores are not comparable with v4.x. Any v4.3 or v4.3.2 figure you find elsewhere, and the pre-September launch figures (Kimi K3 57.1, Grok 4.6 61), cannot sit in the same column.
  • Effort moves the score as much as the model does. Sonnet 5.5 is 54 at max and 41 at high; GPT-6.1 Terra is 51 at max and 38 at its default medium. Compare at the effort you will serve.
  • The top entry may be unavailable. Carbon leads v5.0.1 but has no model ID, and Muse Wonder Full is sold only inside a subscription product; compare against the best model you can actually call.
  • "With fallback" means a routed system. Claude scores include safeguard fallbacks to Opus 4.8 or Opus 5 when a classifier intervened; pin fallback state the way you pin harness and effort.
  • Name the effort and the runner. Astra is 58.18% at max on the leaderboard but 57.9% at high in OpenAI's figure; Fable 5.1 is 57.88% on the leaderboard but 55.8% in Anthropic's own run. The runner can even flip the order: Anthropic's runs put Sonnet 5.5 ahead of Opus 5.5 (70.6% vs 66.4%), and the leaderboard puts it behind (61.82% vs 64.85%).
  • Saturated sets stop discriminating. SWE-bench Verified, the public SWE-Bench Pro split, and GPQA Diamond no longer separate frontier models; use private or held-out sets. See Benchmarks and Leaderboards.

Task-Specific Recommendations (November 2026)

TaskRecommended ModelsWhy
Autonomous Coding AgentClaude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol; GPT-6.1 Terra for cheaper fleetsThe two Claude models lead the Terminal-Bench 4.0 leaderboard (64.85% and 61.82%) and Anthropic's own runs, at $2-$4 input; defaults in Claude Code and Codex; Terra at half the price if it holds up at your serving effort
Complex ReasoningClaude Fable 5.1, GPT-6 Astra (xhigh/max), Claude Opus 5.5 at high effortCeiling models; Opus 5.5 first, escalate when evals fall short
Agentic Computer UseClaude Opus 5.5 / Sonnet 5.5 with computer_toolset_20260801, GPT-6 Astra, Gemini 3.8 FlashOSWorld 2.1 partial 81.8% / 80.1% (Anthropic) and 72.6% for Astra on OpenAI's offline set (OpenAI); Gemini's recommended computer-use model, with the Computer Use tool GA since October 30. Claude Haiku 5.5 scores 72.4% partial on Anthropic's offline subset, but its 24.4% GUI prompt-injection success rate argues for tight isolation
High-Volume APIGPT-6 Luna, Claude Haiku 5.5 (prompts under 100K), Gemini 3.8 Flash, DeepSeek V4.1-Flash (off-peak)Lowest cost per token in class; budget Gemini on its 2027 price and keep Haiku 5.5 under its cliff
Long Context RAGClaude Sonnet 5.5 / Opus 5.5 (flat to 1M), Gemini 3.8 FlashNo long-context surcharge on Claude except Haiku 5.5, which bills 5x above 100K; keep OpenAI prompts under 272K or price the cliff
Ultra-Long ContextLlama 4 Scout (10M)Industry-leading 10M context; open weights
Multimodal Real-timeGPT-Live-1, Gemini 3.8 LiveFull-duplex voice with delegated reasoning, or native audio with async tools
Private ProductionGLM-5.3-Flash, Qwen3.8-27B, MiMo-V2.6-Flash, IBM Granite 4.2Permissive licenses; single-node to single-GPU footprints
Open-source CodingKimi K3.1 Pro, GLM-5.3, MiMo-V2.6-Pro, Kimi K3 (check the license)Top open scores on AA and SWE-Bench Pro v2 private; K3.1 Pro weights since November 4
Gated Cyber or Bio WorkClaude Mythos 5.1, GPT-5.6-Cyber, Gemini 3.8 Flash Cyber, Gemini 4 Carbon (when access opens), GPT-6 Rosalind and GPT-RosalindCapability-tiered access programs with identity verification
Creative/ChatGPT-6 Sol, Claude Sonnet 5.5Strong conversation quality and instruction following at $2/$10

Interview Questions

Q: How would you select a model for a production RAG system?

Strong answer: I evaluate across these dimensions:

1. Quality requirements:

  • Test on representative queries from the actual domain
  • Measure answer correctness, hallucination rate, citation accuracy

2. Cost analysis:

text
Monthly cost = requests/day × 30 × (input_tokens × input_rate + output_tokens × output_rate)

Always calculate for the top 2-3 candidates, at the price that applies next quarter (Gemini 4 Argon doubles on December 1, Gemini 3.8 Flash on January 1, 2027), with your measured cache hit rate, prompt-size distribution, and tokens per task.

3. Latency requirements:

  • Tight TTFT budgets: GPT-6 Luna, Claude Haiku 5.5 at low effort, Gemini 3.8 Flash at low thinking; measure TTFT yourself
  • If quality is paramount: accept multi-second latency with Claude Opus 5.5 or GPT-6.1 Sol, or buy speed with Opus 5.5 fast mode or OpenAI Fast

4. Operational requirements:

  • Self-hosting: GLM-5.3-Flash, Qwen3.8-27B, or Llama 4 Scout/Maverick, after a license check
  • Compliance / data residency: ZDR-eligible models, residency-pinned endpoints (~10% premium), or self-hosted

5. Practical selection:

  • Start with Claude Sonnet 5.5 or GPT-6 Sol for prototyping
  • A/B test GPT-6 Luna, Claude Haiku 5.5, or Gemini 3.8 Flash for the 80% of easy queries (cost)
  • Keep Opus 5.5 on hard queries via semantic routing, and keep prompts under 272K on OpenAI models and under 100K on Haiku 5.5 (retrieved context counts, cached or not)

Q: Explain the tradeoffs between proprietary and open source models.

Strong answer:

FactorProprietary (OpenAI, Anthropic, Google)Open Weights (GLM, Qwen, DeepSeek, MiMo)
QualityHigher at the top end13 to 26 points behind on AA v5.0.1 (46 or 33 vs 59), depending on whether Kimi K3.1 Pro counts as open
CostPer-token pricing; $0.10/$0.50 at the low endCompute + ops; cheap only at high utilization
ControlLimitedFull
PrivacyData goes to provider (ZDR available on many models)Stays on-prem
UpdatesAutomatic, sometimes under the same model IDManual; you pin the weights
CustomizationLimited fine-tuning (OpenAI stops new fine-tuning jobs January 6, 2027, and GPT-4o and GPT-4.1 fine-tunes shut down January 14, 2027)Full fine-tuning
LicenseTerms of serviceRanges from MIT/Apache to MaaS-gated custom licenses
Ops overheadNoneSignificant

Key insight (November 2026): The gap is real, its size depends on the classification rule, and the license is the other big variable. On AA v5.0.1 the best model AA labels open (DeepSeek V4.1-Flash-1017, 33) trails the best closed model you can call (Claude Opus 5.5, 59) by 26 points; Kimi K3.1 Pro, whose weights shipped November 4 but which AA still labels proprietary, narrows that to 13 (46). Always state the index version and the rule. The best US open model (Muse Spark Open, 25) trails DeepSeek by 8. Meanwhile GPT-6 Luna and Claude Haiku 5.5 at $0.10 / $0.50 undercut most hosted open models, so "open is cheaper" only holds at high utilization on your own GPUs. Choose open weights for control, residency, and fine-tuning, and read the license first: Mistral Large 4 gates every MaaS business, and Kimi K3 and K3.1 Pro, Muse Spark Open, Step 5 Preview, MiniMax-M3.1-Flash, Qwen Community 1.0, Mistral Medium 3.5, and GLM-5.3 all gate some commercial uses.

Q: How do reasoning controls differ between GPT-6 and the Claude 5.5 generation, and what breaks when you route between them?

Strong answer: Both vendors expose reasoning as an effort dial rather than a toggle, but the contracts differ:

  • OpenAI GPT-6 family: effort low to max (GPT-6.1 Sol defaults to medium; Astra has no none, while GPT-6.1 Terra runs from none to max). Reasoning tokens are billed but not returned. "Pro" is reasoning.mode: "pro" on the GPT-5.6 and GPT-6 families rather than a separate model (GPT-5.4 Pro and GPT-5.5 Pro remain at $30/$180). Astra drops temperature, top_p, and logprobs, and its tool calling requires the Responses API.
  • Claude Opus 5.5 / Sonnet 5.5 / Fable 5.1: adaptive thinking is always on (Sonnet 5.5's floor is between_tools; disabled returns a 400). Default effort differs by model: Opus 5.5 medium, Sonnet 5.5 and Fable 5.1 high on the API. Thinking, including the text between tool calls, arrives in thinking content blocks that are empty at the default display: "omitted" (the display: "updates" beta returns the text). Blocks are bound to the producing model and conversation. Under API version 2026-10-21 any temperature, top_p, or top_k returns a 400 on every Claude model.
  • What breaks in a router: replaying a Claude conversation to another model silently drops its reasoning blocks; editing anything before a thinking block (system prompt, tools, earlier turns, even the image bytes behind a URL) returns a 400 for accounts created on or after August 31, 2026; forced tool_choice returns a 400 on Opus 5.5, Sonnet 5.5, and Fable/Mythos 5.1, while Haiku 5.5 accepts it but skips thinking on that turn; and an unpinned Opus 5 to 5.5 upgrade quietly moves effort from high to medium.

Production choice: Pin effort per route, keep conversation history append-only, use auto tool choice with strict schemas, and rerun evals at the effort you will actually serve. Log stop_reason, effort, and the served model on every call; visible reasoning is no longer something you can count on from either vendor.

Q: You want to launch a hosted coding assistant on an open-weight model. How do you choose one?

Strong answer: I screen in this order, because the first gate eliminates more candidates than benchmarks do:

  1. License triggers: a hosted coding assistant is exactly what gated licenses target. Mistral Large 4 and Qwen3.8-Flash-Next (Qwen Community 1.0) need a separate license from day one, Muse Spark Open and Step 5 Preview need a separate agreement for MaaS once revenue passes US$20M, Kimi K3 and K3.1 Pro need a separate agreement if the company runs a MaaS business and its total revenue (with affiliates) passes US$20M over 12 months (K3.1 Pro's license defines MaaS as giving third parties meaningful control over inference or fine-tuning, so a coding assistant whose model is embedded in its own features and harness may fall outside it; have counsel read the definition), MiniMax-M3.1-Flash needs at least a one-time notice for any commercial use and written authorization above US$20M yearly revenue, Mistral Medium 3.5 withdraws all rights above US$20M monthly revenue, and GLM-5.3's security review only bites above US$10B. MiMo-V2.6, GLM-5.3-Flash, DeepSeek V4.1-Flash, Ling-3.1-flash, Beam 1.0, Tencent Hy4 preview (preview checkpoint only), and Qwen3.8-27B are MIT or Apache.
  2. Quality on like-for-like scales and on my tasks: AA v5.0.1 (Kimi K3.1 Pro 46, DeepSeek V4.1-Flash-1017 33, GLM-5.3 32, MiMo-V2.6-Pro and Ling-3.1-flash-1019 30; GLM-5.3-Flash 28 on v5.0), at the effort I will serve, and SWE-Bench Pro v2 private (GLM-5.3 211/272, Kimi K3 214/272), then my own eval on real repositories.
  3. Serving cost by architecture: KV bytes per token differ by orders of magnitude across these models, DeepSeek V4.1-Flash activates 8B parameters at prefill but 16B at decode, and several checkpoints ship their own speculative-decoding heads. Size from the model card, not a generic formula.
  4. Supply chain: mirror the exact weights and LICENSE files, watch for API-only refreshes that leave the open checkpoint behind (Qwen3.8-Max-0902), and treat post-trained derivatives as inheriting their base license.
  5. The closed baseline: compare total cost against GPT-6 Luna, Claude Haiku 5.5 (under 100K-token prompts), GPT-6.1 Terra, or Gemini 3.8 Flash. If my utilization is low, the hosted closed model often wins on cost.

Q: A vendor says its new model beats your current one by 10 points. What do you check before switching?

Strong answer: Six things. Some move a number by a point or two; the harness and index-version effects alone have moved headline numbers by tens of points this year:

  • Runner: vendor run or public leaderboard? On Terminal-Bench 4.0, Anthropic's runs put Sonnet 5.5 ahead of Opus 5.5 (70.6% vs 66.4%), while the tbench.ai leaderboard puts it behind (61.82% vs 64.85%).
  • Effort: Sonnet 5.5 is 54 at max and 41 at high on AA v5.0, a 13-point swing from one setting, and GPT-6.1 Terra is 51 at max but 38 at its default medium; compare at the effort you will pay for.
  • Harness: Astra scores 62.7% on ARC-AGI-3 with ARC Prize's Standard harness and 99.9% with ARC Prize's Provider Adapter harness (which uses OpenAI's own context-management features), same model and test set.
  • Fallback state: Claude scores "with fallback" belong to a router plus several models.
  • Index version and saturation: Artificial Analysis re-baselined twice in two months (v4.3 in September, v5.0 in October) and then regraded every model in v5.0.1 (November 3), and each time scores moved, so a "10-point gain" across versions means nothing. Opus 5 scores 99.4% on the public SWE-Bench Pro split but 81.6% on the private one.
  • Cost per task: tokens per task differ by about 3x between models of similar quality.

Then I run my own eval at the serving effort, on held-out tasks, with fallback state pinned, and I rerun it after vendor-side changes (GPT-6 Sol's September 25 image fix changed results under the same ID).


Release History (2026)

Earlier monthly headlines, kept for context. Each describes the landscape as it stood at the time; the sections above describe it as of November 6, 2026.

October 2026 (mid-August to October 1). Anthropic, OpenAI, Google, SpaceXAI, and Meta all shipped new models in September, and the story was price as much as capability. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 (September 1, same $10/$50 with cache reads cut to 0.025x), Claude Opus 5.5 (September 22, the first Opus price cut since Opus 4.5, to $4/$20, which Anthropic made its "start here" default), and Claude Sonnet 5.5 (September 28, $2/$10). OpenAI shipped GPT-6 Astra (September 3, $10/$50, a Daybreak-first staged rollout and the first OpenAI model rated Critical for cybersecurity), GPT-6 Sol and Luna (September 22, $2/$10 and $0.10/$0.50, about half the GPT-5.6 prices they replaced, with no Terra tier until GPT-6.1 Terra on October 26), and GPT-6.1 Sol (September 29, same $2/$10 with 0.05x cache reads). Google shipped Gemini 3.8 Flash (September 2, its third Flash in six weeks, half price until January 1, 2027) plus a Fairwind-gated 3.8 Flash Cyber, then on September 30 announced Gemini 4 Argon, rolling out to cyber defenders first (GA followed on October 13). SpaceXAI shipped Grok 4.7 (September 21) on a new, larger base at an unchanged $2/$6, and Meta shipped Muse Spark 1.3 (September 2). The practical consequences: the mid tier converged on $2/$10 (Sonnet 5.5, GPT-6 Sol, GPT-6.1 Sol), cache-read multipliers began to differ by model (0.1x, 0.05x, 0.025x), OpenAI billed the whole request at long-context rates above 272K input, and the September Claude models rejected forced tool_choice and bound thinking blocks to the producing model and conversation, which broke naive routers. Benchmarks reset too: Artificial Analysis re-baselined its Intelligence Index (v4.3, September 7, itself replaced by v5.0 on October 27), Terminal-Bench 4.0 replaced 3.0, the public SWE-Bench Pro split saturated, and Claude scores began carrying a "with fallback" label. On open weights, Z.ai shipped GLM-5.3's weights on schedule (August 28) under an MIT-plus license, Xiaomi's MiMo-V2.6-Pro (September 21, MIT) took the top open-weight spot, DeepSeek V4.1-Flash (September 10, MIT) cut Flash prices less than a month after the August 16 increase, and Qwen's new Community License required a separate license for any MaaS or coding/office-assistant business. Retirements: the Assistants API shut down August 26, gpt-5.4-cyber was shut down October 1 on 20 days' notice, the original gpt-4o-2024-05-13, o1, o3-mini, and o4-mini snapshots were scheduled to leave the OpenAI API October 23 (they did), and Claude Sonnet 4.5 was deprecated for November 30.

August 2026. Most of the month's releases were post-training refreshes or derivatives rather than new pretraining runs, and no US frontier lab shipped a new flagship base model. The action moved to prices, licenses, and access control. The clear exception was Alibaba's Qwen3.8-Max, a genuinely new 2.4T base whose weights landed August 12. Claude Sonnet 5's introductory $2/$10 per 1M became the permanent price on August 10, and the scheduled September 1 rise to $3/$15 was canceled, making Sonnet 5 permanently cheaper than the Sonnet 4.6 it replaced. Running the other way, DeepSeek raised V4 prices 3x to 12x effective August 16 at 16:00 UTC, moving to peak and off-peak billing (off-peak is half of peak, and peak covers only 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays). That ended its run as the unambiguous cheap option. OpenAI shipped GPT-5.6-Cyber (August 10, $12.50/$75 per 1M) behind a new two-tier Daybreak Blue and Daybreak Red access program, at the time the most concrete production example of capability-tiered gating. Google released Gemini 3.7 Flash (August 13) at half price through year-end, and SpaceXAI released Grok 4.6 (August 12) with a 500K context. On open weights the licensing picture split: Alibaba's Qwen3.8-Max (August 12, 2.4T/95B active) shipped under a bespoke commercially gated license while its Qwen3.8-27B sibling (August 14) shipped plain Apache 2.0, Meta returned to open weights with Muse Glimmer (August 10, 30B, Apache 2.0), and Z.ai withheld GLM-5.3's weights pending safety evaluation after cyber capability grew faster than expected (the weights then shipped on schedule, August 28). Claude Opus 4.1 retired on August 5 on the Claude API, closing out the last $15/$75 Opus tier. Most benchmark figures in that month's launch posts were vendor-reported.

July 2026. Anthropic shipped a full generation refresh: Claude Sonnet 5 (June 30, claude-sonnet-5, which became the default everywhere, at an introductory $2/$10 per 1M made permanent on August 10) and Claude Opus 5 (July 24, claude-opus-5, unchanged $5/$25 with an optional Fast mode at $10/$50, about 2.5x faster). Claude Fable 5 was restored globally July 1 after the export-control suspension, with a new jailbreak-specific cybersecurity classifier; Mythos 5 returned only to roughly 100 US critical-infrastructure organizations via Project Glasswing. GPT-5.6 (Sol, Terra, Luna) reached general availability July 9, and on July 30 OpenAI cut Luna 80% to $0.20/$1.20 and Terra 20% to $2/$12 (Sol stayed at $5/$30 until a promotional cut to $4/$20 on August 21). The open-weight frontier had its strongest month to that point: Moonshot Kimi K3 (July 16, weights July 27) became the largest open-weight model to date at 2.8T total / 104B active with a 1M context, and Thinking Machines Lab debuted Inkling (July 15, 975B / 41B active, open weights), then the leading US open-weights model. Google shipped Gemini 3.6 Flash, 3.5 Flash-Lite, and a government-gated 3.5 Flash Cyber (July 21) while delaying Gemini 3.5 Pro, which never shipped. Meta Muse Spark 1.1 (July 9) arrived with Meta's first paid self-serve model API ($1.25/$4.25 per 1M). Black Forest Labs announced FLUX 3 (July 23), a unified image, video, audio, and action model, in gated early access. Most benchmark figures in these launches were vendor-reported.

June 2026. Anthropic released Claude Fable 5 (June 9, claude-fable-5, $10/$50 per 1M, 1M context), then its most capable widely released model: a Mythos-class model made safe for general availability, with an Opus 4.8 fallback safeguard on sensitive topics. Claude Mythos 5 shipped the same day as the unrestricted variant for Project Glasswing partners, succeeding Mythos Preview at less than half its price.

June 10-26, 2026. A dense second wave of June launches followed. Google DeepMind DiffusionGemma (June 10, Apache 2.0), Google DeepMind's first open-weight text-diffusion model, is a 26B Mixture-of-Experts (~4B active) that denoises blocks of tokens in parallel for roughly 4x faster generation on a single H100, trading some quality versus standard Gemma 4. Gemini 3.5 Live Translate (June 9 per launch coverage; Google's changelog had no entry for it, and it was listed as gemini-3.5-live-translate-preview) added real-time speech-to-speech translation across 70+ languages in public preview via the Gemini Live API and AI Studio. Cohere North Mini Code 1.0 (June 9, Apache 2.0), Cohere's first open coding model, is a 30B / 3B-active MoE that runs on one H100. Moonshot Kimi K2.7 Code (June 12, Modified MIT) tuned K2.6 for long-horizon software work (1T / 32B-active MoE, roughly 30% fewer thinking tokens). Z.ai GLM-5.2 (coding-plan access June 13, open weights under MIT June 16-17), a 744B / 40B-active MoE with a 1M context, reported SWE-Bench Pro 62.1, ahead of GPT-5.5 on that benchmark, at roughly $1.40 / $4.40 per 1M. xAI Grok Imagine Video 1.5 reached general availability June 16 (image-to-video with synchronized audio, $0.080 per second of video), and Grok 4.3 arrived on Amazon Bedrock June 15 ($1.25 / $2.50 per 1M, xAI's first model there). Alibaba's Qwen Cloud changelog listed a June snapshot adding vision to Qwen 3.7-Max (text-only at its May launch), though some independent coverage attributed that update to Qwen 3.7-Plus instead. Separately, on June 12 Anthropic suspended access to Claude Fable 5 and Claude Mythos 5 following a US export-control directive, with Mythos 5 later cleared for a limited set of US institutions. Then on June 26, OpenAI previewed GPT-5.6 (Sol, Terra, and Luna) in a limited release to a small set of US-government-approved partners over dual-use cybersecurity concerns, echoing the Anthropic restriction; OpenAI claimed a new Terminal-Bench 2.1 record for Sol and pitched Terra at GPT-5.5-level quality for about half the cost. Most coding scores in these launches were vendor-reported.

May 2026. Anthropic Claude Opus 4.8 (May 28, same $5/$25 price as Opus 4.7; Dynamic Workflows research preview with hundreds of parallel subagents; fast mode at $10/$50, 3x cheaper than the Opus 4.7 fast mode); OpenAI GPT-5.5 (April 23) and GPT-5.5 Instant (May 5, default in ChatGPT); Claude Opus 4.7 (April 16, GA on Bedrock/Vertex/Foundry); Google Gemma 4 (April 2, Apache 2.0) and Gemini 3.2 Flash (quiet rollout May 5); DeepSeek V4 Pro and V4 Flash (April 24; the 75% V4 Pro discount was made permanent May 22, with a $0.435/$0.87 per 1M list price from June 1 that lasted until the August 16 repricing); Moonshot Kimi K2.6 (April 20, 1T MoE / 32B active); Alibaba Qwen 3.6 Plus / 3.6-35B-A3B / 3.6 Max-Preview; Mistral Medium 3.5 (April 29, unified chat/reasoning/coding/vision); Meta Muse Spark (April 8, first closed-weight Meta model). Meta also paused the Llama 4 Behemoth release amid capability concerns; it never shipped. Published SWE-bench Verified leaders before the Fable 5 launch: Claude Mythos Preview 93.9%, GPT-5.5 88.7%, Claude Opus 4.8 88.6%; ARC-AGI-2 leader: GPT-5.5 at 85.0%. Anthropic described Fable 5 as state of the art on nearly all tested benchmarks but put no standard numeric scores in its launch post.


References


Next: Capability Assessment

From the AI System Design Guide by Om Bharatiya · MIT License ·view source