LLMs, GenAI, RAG, agents, evaluation · insight-0664
Pin dated model versions and schedule migrations before retirement dates
- Evidence type
- documented fact
- Difficulty
- easy
- Roles
- ML engineer, Applied AI
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
Production calls a moving model alias, or a pinned model that the provider has deprecated.
Action
Pin explicit model versions, track the provider's deprecation schedule in your risk register, and run the golden set on the candidate replacement well before the retirement date.
Reason
Providers retire models on a schedule; Anthropic documents active, legacy, deprecated and retired states, says requests to retired models fail, and advises testing replacements well before retirement.
Boundary
Every application on hosted models; self-hosted weights avoid retirement but not security patches.
Symptom or risk
Behaviour changes without a release on your side, or requests start failing on the retirement date with no tested replacement.
Exceptions
Low-risk internal tools where automatic upgrades are acceptable and monitored.
Trade-offs
Pinning delays improvements and requires periodic migration effort.
Pitfalls
Migrating by swapping the model id without re-running evaluations, cost estimates and latency tests.
Lessons and concepts
Claims and their sources
Anthropic describes active, legacy, deprecated and retired model states, states that requests to retired models fail, and, as read on 2026-10-11, commits to at least 60 days' notice before retiring publicly released models (check the current page).