Every model your product calls today will be retired, and after that date your requests to it fail. This is for teams running AI features in production. The short answer: pin exact model versions, keep an eval set, keep the model name in config, budget for request changes, and know each tool's notice period before you need it.
Anthropic's model deprecation policy is typical for API vendors: at least 60 days' notice for publicly released models, and after retirement, requests to the model fail. We run production AI pipelines and move them across model versions. Here's what makes that move boring.
1. Pin the exact version in production. Use the dated snapshot ID, not a moving alias. An alias can point at a new model without a deploy on your side. Then output quality shifts and nobody can say why. Keep aliases for prototypes.
2. Keep an eval set you can run in an afternoon. That's fifty to a few hundred real examples with known-good outputs, scored automatically. Without one, a retirement notice starts weeks of meetings. With one, it starts a test run. We cover how to build it in our guide to reading model releases like a buyer.
3. Put the model name in config, not code. Route provider calls through one interface and keep the model ID in configuration. Then a migration is a config change plus an eval run. It isn't a search through the codebase for hard-coded strings.
4. Budget for a code change, not a config flip. A newer model is rarely a drop-in replacement. Defaults change, request settings get renamed or retired, and output that parsed cleanly can come back shaped a little differently. In our own upgrades, swapping the model ID is the smallest part of the job. The request settings and the output handling take the time.
5. Know the notice period for every tool you rent, not just your API. API vendors publish their windows. Packaged tools often don't. On October 2, 2026, GitHub deprecated four models in Copilot, including Claude Opus 4.7, and the post names replacements that admins may need to enable themselves. If your team built a workflow around one of those models, you found out from a changelog. Write down which tools you depend on and how much warning each one actually gives.
Treat a 60-day notice as your deadline to finish, not your signal to start. If your AI feature has no plan for the day its model goes away, that plan is part of our AI implementation work, along with the portability checks we run before committing to a coding agent.