Model deprecation

Providers cannot serve every historical model forever, so older versions move through a lifecycle: deprecation announcement, a migration window measured in months, and final retirement when the endpoint stops responding. Snapshot-pinned versions get dated shutdown timelines precisely so production users can plan.

Migration is rarely a drop-in swap. Successor models differ in behavior, verbosity, and occasionally price structure, so prompts need re-testing and evals need re-running — an operational cost that recurs with every retirement cycle.

Teams that track provider deprecation calendars, keep an eval suite ready, and abstract the model identifier behind configuration handle these transitions in days; teams that hard-code model names across a codebase discover deprecation the expensive way.