Model deprecation

Model deprecation is the lifecycle event where a provider announces and eventually removes an LLM snapshot — production apps must migrate to a newer model or pinned version before the deprecation date or queries start failing.

Every major LLM provider deprecates older snapshots: OpenAI typically gives 6-12 months notice (GPT-3.5 turbo legacy snapshots, instruct-davinci-003), Anthropic gives 6-12 months on Claude snapshots, Google on Gemini. The deprecation cycle: announcement → grace period → soft-fail (warnings) → hard-fail (404). Production playbook: monitor provider changelog, track which app code paths use which pinned model, eval the candidate replacement against your golden set, migrate with a feature flag, decommission old pin. Skipping any step risks production fires. By 2026 model deprecation is a quarterly recurring task for any production AI team. [[Model version pinning]] and a clean migration story prevent fire drills.

When to use model deprecation

• Production AI teams must track deprecations.

Common mistakes

• Not monitoring provider changelog — surprise hard-fail.
• Migrating without re-evaluating — quality regression slips into prod.

FAQ

Related terms

• Model version pinning — Model version pinning is the production practice of locking your app to a specific model snapshot (e.g., `claude-sonnet-4-6-20250101`) instead of an alias (`claude-sonnet-4-6`) — protects against silent behavior changes when the provider rolls a new snapshot.
• Model registry — A model registry is the versioned store for trained model artifacts + metadata — the source of truth for which model version is in staging vs prod, what its eval scores were, and how to roll back. MLflow, W&B, Sagemaker, Vertex AI all ship one in 2026.

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/model-deprecation.md.