The model behind an app can change without the app changing

Start here

A workflow that reliably formatted a spreadsheet or drafted an email a certain way for months can suddenly behave differently, with no visible update to the app itself. Often the interface has not changed; the model behind it has, because a vendor retired the version it was quietly running on. Anthropic's model deprecations page states this lifecycle in its own terms: models move from active to deprecated, where a recommended replacement and a retirement date are assigned, and finally to retired, where requests to retired models will fail outright.

What the documents say

All three vendors commit, in their own documentation, to advance notice, though the length differs. Anthropic's page states it provides at least 60 days' notice before retirement for publicly released models, and its deprecation history lists past cases, such as notifying developers on 14 April 2026 that Claude Sonnet 4 and Claude Opus 4 would retire by 15 June 2026. OpenAI's deprecations page commits to longer windows for its most-used models, six months for generally available models against as little as two weeks for preview models, and lists concrete shutdowns, including legacy GPT snapshots retiring on 23 October 2026. Google's Gemini API release notes record deprecation as dated entries rather than one policy statement: an entry dated 15 June 2026 states named image-generation models would be shut down on 17 August 2026, roughly two months' notice. None of the three documents a way to keep a retired model running past its stated date.

Check this

A person relying on a specific model version, rather than whatever the app currently uses, can check the vendor's deprecation page for that exact model string against the retirement date listed there. Anthropic's documentation also describes a usage audit a customer can run directly: exporting API usage from its console and reviewing it by model, to spot lingering calls to a model already marked deprecated before its hard retirement date arrives.

What holds and what fails

The claim that a workflow can silently break because of a vendor-side model change, not a mistake in the workflow itself, holds according to every policy examined here, all of which describe retirement as routine rather than exceptional. It fails as a complete explanation whenever a product shields its users from model identifiers altogether, in which case a person has no string to check against these pages, only a date when behavior seemed to shift. Assuming a working prompt or setting will keep working indefinitely is the assumption these pages exist to correct.

• Note the specific model name or version string behind any workflow that must keep behaving consistently.
• Check the vendor's deprecation page periodically for that model rather than waiting for a workflow to break.
• Treat a sudden, unexplained change in a familiar tool's output as a possible model swap, not only a personal mistake.

A retirement date on a vendor's page is not fine print; it is a scheduled event with a real deadline. Reading the specific page for the model actually in use is the only way to see that deadline before it arrives.