Skip to content
anythingai.space
Deployment & Deliveryintermediate11 min · 9 steps

Model Versioning

Pin the exact chat model snapshot, embedding model and fine-tuned weights per environment, record them on every request, and treat any change as a gated release.

tool-call Δ-3 pts
cost/turn Δ8%
provider moved itrag-apimodel: chat-latestAliaschat-latestSnapshot AretiringSnapshot Bsilently swapped in
requesterrordata
metrics are illustrative
1 / 9
ProblemStep 1 of 9

The model changed and nobody deployed anything

The assistant calls the provider's undated alias. One morning the provider points that alias at a newer version. Answers get longer, one tool's arguments come back in a different format, refusal behaviour shifts, latency and cost move.

Nothing in the release history explains it — because nothing was released. And older versions are retired on a schedule, so even a pinned model must eventually move.

Watch outA floating alias is an unreviewed, ungated model upgrade on the provider's schedule.

Lesson 17 of 20 in Deployment & Delivery