OpenAI gives about 2.9 times the notice Anthropic does before retiring a model
On their own deprecation pages, read on 26 October 2026, OpenAI states a median of 182 days between telling developers and shutting a model down, and Anthropic 63 days. Google's page gives shutdown dates but not notification dates. Below: the notice for each announcement, what is scheduled to retire in the next 150 days, and whether the recommended replacement costs less.
In this article
What the providers' own pages say
We read the deprecation pages of OpenAI, Anthropic and Google on 26 October 2026, saved copies with hashes, and parsed their tables. Notice is the gap between the date a provider says it notified developers and the shutdown date.
Across 11 distinct announcements on OpenAI's deprecations page (range 20 to 184 days), shutdown date minus the notification date it states.
Across 10 distinct announcements on Anthropic's deprecations page (range 60 to 189 days). Its recent notices are all 60 to 62 days.
Google's page lists release and shutdown dates but not when developers were told, so we compute no notice figure for it.
Retirements and earliest-possible retirements stated by the three providers between 26 October 2026 and 25 March 2027.
How much notice, announcement by announcement
OpenAI's median notice is 182 days and Anthropic's is 63, about 2.9 times as long. The charts show the most recent announcements for each.
OpenAI, days between notification and shutdown (by notification date)
Anthropic, days between notification and retirement (by notification date)
What is scheduled to go next
Stated dates within the next 150 days, soonest first. "Earliest possible" rows are Anthropic's tentative floors for active models, not announcements.
| Date | Provider | Model | Recommended replacement | What the page says |
|---|---|---|---|---|
| 17 November 2026 | gemini-3.1-flash-tts-preview | gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts | Stated date | |
| 17 November 2026 | gemini-3.1-flash-live-preview | gemini-3.8-live | Stated date | |
| 17 November 2026 | gemini-2.5-flash-native-audio-preview-12-2025 | gemini-3.8-live | Stated date | |
| 17 November 2026 | gemini-2.5-flash-preview-tts | gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts | Stated date | |
| 17 November 2026 | gemini-2.5-pro-preview-tts | gemini-3.8-flash-tts or gemini-3.8-flash-lite-tts | Stated date | |
| 24 November 2026 | Anthropic | claude-opus-4-5-20251101 | none stated | Earliest possible |
| 30 November 2026 | Anthropic | claude-sonnet-4-5-20250929 | claude-sonnet-5-5 | Stated date |
| 1 December 2026 | OpenAI | gpt-image-1-mini | gpt-image-2.5-sunburst or gpt-image-2.5-flare | Stated date |
| 1 December 2026 | OpenAI | gpt-image-1.5 | gpt-image-1.5-2025-12-16 | gpt-image-2.5-sunburst or gpt-image-2.5-flare | Stated date |
| 1 December 2026 | OpenAI | chatgpt-image-latest | gpt-image-2.5-sunburst or gpt-image-2.5-flare | Stated date |
| 11 December 2026 | OpenAI | gpt-5 | gpt-5-2025-08-07 | gpt-5.6-sol | Stated date |
| 11 December 2026 | OpenAI | gpt-5-mini | gpt-5-mini-2025-08-07 | gpt-5.6-terra | Stated date |
| 11 December 2026 | OpenAI | gpt-5-nano | gpt-5-nano-2025-08-07 | gpt-5.6-luna | Stated date |
| 11 December 2026 | OpenAI | gpt-5-pro | gpt-5-pro-2025-10-06 | gpt-5.6-sol ( reasoning.mode: pro ) | Stated date |
| 11 December 2026 | OpenAI | o3 | o3-2025-04-16 | gpt-5.6-sol | Stated date |
| 11 December 2026 | OpenAI | o3-pro | o3-pro-2025-06-10 | gpt-5.6-sol ( reasoning.mode: pro ) | Stated date |
| 6 January 2027 | OpenAI | tts-1 | gpt-realtime-2.1-mini | Stated date |
| 6 January 2027 | OpenAI | tts-1-hd | gpt-realtime-2.1-mini | Stated date |
41 rows in the window; the first 18 are shown. The full list is in retirements.json in the dataset, and every date is in a subscribable calendar (.ics download); earliest-possible rows are marked tentative there.
Is the replacement cheaper?
For the Anthropic retirements where both models are in our saved pricing page, the list price change from the retired model to the recommended replacement:
| Retired model | Recommended replacement | List price change | Estimate with the new tokenizer |
|---|---|---|---|
| claude-sonnet-4-5-20250929 | claude-sonnet-5-5 | -33.3% | -13.3% |
| claude-opus-4-1-20250805 | claude-opus-4-8 | -66.7% | -56.7% |
| claude-3-5-haiku-20241022 | claude-haiku-4-5-20251001 | +25% | not applicable |
Derived by us from list prices (input plus output, per million tokens). The tokenizer estimate multiplies the replacement's price by 1.3, from Anthropic's statement that Claude 4.7 and later produce about 30% more tokens; it is an estimate, not a measured bill. OpenAI and Google replacements are not priced here because the retired models are not in our saved pricing pages.
A name that is retired can also keep working by routing elsewhere. Google's page says gemini-3.7-flash and gemini-3.5-flash automatically route API requests to newer models.
What we saw in our own probes
Separately from what the providers state, we probe whether the models used in our own studies still answer. In the reliability study we published (lucreya.com), 5 of the 9 models probed had been retired by their provider, meaning the provider returns model_not_found. That is our observation on one roster, not a market figure.
How we know, and what we did not verify
Dates, replacements and notification dates come from the HTML tables of the pages below. The build stops if a table is missing.
| Page | Date | SHA-256 (first 16) |
|---|---|---|
| platform.claude.com/docs/en/about-claude/model-deprecations | 26 October 2026 | 302f8bbb922bccea |
| developers.openai.com/api/docs/deprecations | 26 October 2026 | 48c8cd40f4641e5b |
| ai.google.dev/gemini-api/docs/deprecations | 26 October 2026 | 0f89496657fafb81 |
- Only Anthropic, OpenAI and Google. xAI, DeepSeek and Mistral were not read.
- Notice figures exist only where a page states the notification date.
- We have not asked any provider to comment before publication. A provider or reader who finds a wrong figure can tell us; we correct it and date the correction.
Part of the series: price calendar, OpenAI pricing, Gemini API pricing, Groq: stated shutdown dates against what we observed.
FAQ
How much notice does OpenAI give before removing a model?
On its deprecations page, read on 26 October 2026, OpenAI states a notification date and a shutdown date for each announcement. Across 11 distinct announcements the median gap is 182 days, with a range of 20 to 184 days. The 20-day case is a single cyber model; the rest are 92 days or more.
How much notice does Anthropic give?
Anthropic's page states when developers were notified and the retirement date. Across 10 distinct announcements the median is 63 days (range 60 to 189), and its five most recent announcements all gave 60 to 62 days.
Does a retired model name always stop working?
Not always. Google's page says gemini-3.7-flash and gemini-3.5-flash are not shutting down and automatically route API requests to newer models (gemini-3.8-flash and gemini-3.6-flash). DeepSeek's page says the retired deepseek-v4-flash name is served by a newer model and billed at the Flash price. A name that keeps working can change what you get and what you pay.
Is the replacement cheaper?
Sometimes. On Anthropic's pages, Claude Sonnet 4.5's recommended replacement, Sonnet 5.5, lists at a third less per token, and Opus 4.1's replacement, Opus 4.8, at two thirds less. Haiku 3.5's replacement, Haiku 4.5, lists 25% higher. Anthropic also says Claude 4.7 and later use a tokenizer that produces about 30% more tokens, so we show an estimate with that adjustment. We computed this only for Anthropic.
What is "not sooner than"?
For active models Anthropic lists a tentative earliest retirement date. It is a floor, not an announcement: the model is not scheduled to retire on that date.