Two late-October API shutdowns require more than updating a model dropdown. OpenAI's deprecation record says a group of legacy endpoints shut down on October 23, 2026. Google's Veo guide confirms three Veo 3.1 preview endpoints in the Gemini API shut down on October 22. These are effective retirement dates, rather than newly announced deadlines.
OpenAI's legacy shutdown covers several families
The OpenAI record traces its notice to April 22. Its October table includes GPT-3.5 Turbo, older GPT-4 variants, GPT-4.1 nano, an older GPT-4o snapshot, GPT Image 1, and o1, o3-mini and o4-mini entries. It also says fine-tuned versions follow their underlying base model's shutdown schedule.
The listed replacements vary by entry. For example, the table maps the old GPT-3.5 Turbo entry to GPT-5.6 Terra, GPT Image 1 to Sunburst or Flare, and o1-pro to GPT-5.6 Sol with pro reasoning mode. An application using several model types should therefore follow each row's migration path instead of assuming one new identifier covers everything.
The Veo shutdown is scoped to an API surface
Google identifies veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview as retired in the Gemini API. Its guide points developers to Gemini Omni Flash or GA models through the Gemini Enterprise Agent Platform.
The general Gemini deprecation table describes dates as earliest possible shutdowns, but the Veo guide now explicitly confirms this shutdown occurred. The notice concerns these preview endpoints on this surface; it should not be expanded into a claim that every Veo offering has ended.
Look for dependencies outside the main request path
A model reference may live in a scheduled batch, an evaluation fixture, an administrative setting or a job saved before retirement. Search both the code and the configuration that actually selects the endpoint. Assign one owner to each active workflow so a stale reference does not survive between deployment surfaces.
A useful inventory distinguishes a configured default from an explicit per-request override. It should also identify any persisted job arguments that can reintroduce a retired model during replay. This matters for a system that appears healthy during a manual test but fails when an older queued task resumes overnight.
Prove the replacement contract
For text and reasoning tasks, compare accepted outputs, tool arguments, latency and total cost using representative cases. Preserve application validation and review controls while changing the provider request. A successful HTTP response is evidence of connectivity, not evidence that the business workflow still reaches the right outcome.
For video generation, exercise submission, status tracking, output retrieval and the application's supported media settings. Confirm that downstream storage and publishing can handle the actual returned artifact. A replacement model can be available while the surrounding adapter still expects a retired endpoint's behavior.
Retirement work is complete when every live caller uses a supported path, queued work has an explicit handling policy, and the resulting product behavior passes its acceptance checks. The October deadlines make that ownership urgent. Give the maintenance record a named owner and a review date, and connect it to the applications affected so the next provider notice reaches the people who can act. They also show why model lifecycle records belong alongside integration maintenance rather than being treated as occasional release-news reading.