An automation that calls a hosted AI model depends on that model staying available, and hosted models do not stay available forever. Providers publish retirement schedules; after the date, requests to the old model fail. If nobody on your side watches those dates, a workflow that ran flawlessly for a year can stop on an ordinary Tuesday.
What the lifecycle looks like
Anthropic’s model deprecations page defines four terms. A model is active when it is fully supported and recommended. It is legacy when it will no longer receive updates and might be deprecated in the future. Deprecated means it still functions but is no longer recommended, with a recommended replacement and an assigned retirement date. Retired means it is no longer available, and requests to retired models fail. The page also warns that deprecated models are likely to be less reliable than active ones.
As of the page’s reading on 2026-11-04, its status table shows an example worth studying. A dated Sonnet 4.5 model was marked deprecated on September 30, 2026, with a tentative retirement date of November 30, 2026 and a named replacement. That is a two-month window. The same page states that the provider gives at least 60 days’ notice before retiring publicly released models, by email to customers with active deployments and in the documentation. Dates and model names change, so check the live page rather than relying on this example.
Why this is an operations problem, not a developer problem
The notice goes to whoever created the account. In a small business that might be a former contractor’s email address, or an inbox nobody reads. The failure that follows is not subtle: the automation simply starts returning errors. If you built a failure path and a pause switch, you contain the damage. If you did not, customers notice first.
A model-lifecycle checklist
| Item | What to record | Who |
|---|---|---|
| Model identifier in use | The exact name each workflow calls | Integrator or builder |
| Account contact | A monitored shared mailbox, not a person’s inbox | Owner |
| Status check date | Monthly look at the provider’s deprecation page | Named person |
| Retirement date | Copy it into the shared calendar the day it appears | Named person |
| Migration test date | At least one month before retirement | Builder |
Find out what you are actually using
The page describes an audit route: in the provider’s console, the usage page can export a CSV broken down by API key and model. Reviewing it shows any place your applications still call a deprecated model, including old experiments you forgot. If a vendor or agency runs the integration for you, ask them to provide the same list for your account and name every workflow behind each entry. Also note that partner platforms such as cloud marketplaces set their own retirement schedules, so a date on one platform may not match another’s.
Parameters retire too
The deprecations page also covers request settings. One example: sampling settings such as temperature, top_p and top_k were removed in an API version dated 2026-10-21, and requests that set them return a 400 error under that version; the recommended replacement is to omit them and guide behavior through prompting. You do not need to understand the parameters. You need to know that a routine software update by a developer, such as moving to a newer API version, can break a request that has worked for years. Ask your integrator to read the provider’s migration guide for the target model before changing versions, and to run your saved test cases afterward. The provider’s migration guides page indexes a separate guide for each current model family, which is the right place to start.
The migration test
- Use the cases you already have. Re-run the representative test set from your original evaluation against the replacement model (see our article on testing with representative cases).
- Compare, don’t assume. The documentation advises thorough testing with the new model well before the retirement date, because a newer model can behave differently: different wording, different length, different refusals.
- Check cost and speed. Replacement models can differ in price and response time. Re-estimate monthly cost before you switch.
- Switch in a controlled way. Move one workflow first, watch the run history for a week, then move the rest.
- Keep the rollback until the old date. Until the old model actually retires, you can usually switch back.
What to write in your operating notes
For every AI-using workflow, keep one line: the model it calls, the date you last confirmed its status, and the retirement date if any. Review the lines monthly alongside run history. This takes ten minutes, and it turns a surprise outage into a scheduled task with weeks of lead time.



