{"slug": "openai-gives-six-months-anthropic-gives-two-google-gives-weeks", "title": "OpenAI gives six months. Anthropic gives two. Google gives weeks.", "summary": "OpenAI's 38 API and model shutdown notices carry a median of 182 days between announcement and switch-off, Anthropic's nine notices a median of 63 days, and Google's 24 dated notices a median of 28 days, according to a study of the three labs' published deprecation records read on September 27, 2026. OpenAI's deprecations page sets a six-month minimum for generally available models and three months for specialized variants, Anthropic's stated minimum is 60 days, and Google's shortest notices included 15 days for three generally available Veo video models. Fifteen notices gave less than 30 days, ten of them for preview, experimental or beta models and thirteen of them from Google, while OpenAI closed its Videos API with no replacement and gave its Assistants API a year.", "body_md": "The Study · Comparison\n\n# OpenAI gives six months. Anthropic gives two. Google gives weeks.\n\nOn September 24 OpenAI’s Sora API reached its shutdown date. OpenAI had announced it six months earlier and named no replacement, so anyone with a video feature built on it had to rebuild that feature somewhere else.\n\nHow long you get before a model or API disappears is decided mostly on the day you pick it: by the lab you build on, and by whether the model’s name says preview. We read every shutdown notice OpenAI, Anthropic and Google have published and measured the gap between each announcement and the switch-off.\n\n## The short version\n\n- **OpenAI gives about six months.** Its 38 notices have a median of 182 days between announcement and shutdown, and 13 of its 15 notices for generally available models gave at least 179. Its[deprecations page](https://developers.openai.com/api/docs/deprecations) sets a six-month minimum for those and three months for specialized variants, and warns that previews can get much less.\n- **Anthropic gives two months, and lately almost exactly two.** Its nine notices have a median of 63 days, and the last four gave 62, 60, 62 and 61 against a[stated minimum of 60](https://platform.claude.com/docs/en/about-claude/model-deprecations) . Every active Claude model also lists a retirement floor, and for the ten we checked it falls exactly one year after launch.\n- **Google gives weeks.** Its 24 dated notices have a median of 28 days, and its four notices for stable models a median of about 53. Three generally available Veo video models got 15. Seven more retirements, among them the Gemini 1.5 models, show up in the[release notes](https://ai.google.dev/gemini-api/docs/changelog) only after the fact.\n- **Short notice lives in previews, and mostly at Google.** Fifteen notices gave less than 30 days. Ten of them were for preview, experimental or beta models, and thirteen came from Google.\n- **A product shutdown costs more than a model swap.** OpenAI gave its Assistants API a year and Agent Builder six months, and it closed the[Videos API](https://developers.openai.com/api/docs/deprecations) with no replacement. Swapping a model changes a line of config; leaving a hosted product means rebuilding what sat on top of it.\n- **Access can close before the shutdown.** OpenAI stopped organizations that had never fine-tuned from starting jobs on the day it announced the change, and on September 18 Google limited Gemini 2.5 to existing users, adding that the models are not deprecated.\n\nEverything below is the evidence, and the full ledger ships with this piece as a spreadsheet you can check line by line.\n\n## How we built the ledger\n\nWe read three primary records on September 27, 2026. OpenAI’s API deprecations page covers June 2022 to September 2026, and Anthropic’s model deprecations page covers September 2024 to June 2026. For Google we used the Gemini API release notes, which its deprecations page names as the place announcements are made; the dated shutdown notices there run from August 2025 to September 2026. Where Google’s deprecations table lists a model, we cross-checked its shutdown date there.\n\nEach row is one announcement date, one shutdown date and one kind of thing retired. Entries that share all three merge, so a batch of fourteen models retired on the same day counts once, while an announcement covering two kinds, or setting two dates, becomes two rows. Notice is the number of calendar days from announcement to shutdown. The kinds are generally available models, previews (including experimental and beta models), OpenAI’s specialized variants such as its chat-latest and Codex models, and APIs, endpoints and products.\n\nA few rules keep the count true to the record. A model listed in two announcements counts once, in its latest one. We left out OpenAI’s June 2023 notice, whose dates were later extended and whose models reappear in later notices. Google retirements with no dated announcement cannot be measured, so they are listed separately below. The labs also email active users, so some customers may have heard before a public page changed; our dates are the public record. Partner platforms such as Amazon Bedrock and Google Cloud run their own schedules, and parameter changes and alias swaps are left out. The full ledger is [a CSV file](https://okaneland.com/features/study/shutdown-ledger/shutdown-ledger.csv), one row per notice, each with its source.\n\n## The ledger at a glance\n\n|  | OpenAI | Anthropic |  | \n|---|---|---|---|\n| Notices measured | 38 | 9 | 24 | \n| Notices in the last 12 months | 18 | 5 | 22 | \n| Median notice, all | 182 days | 63 days | 28 days | \n| Median notice, generally available models | 183 days (15) | 63 days (9) | about 53 days (4) | \n| Median notice, previews | about 152 days (8) | none retired | 28 days (17) | \n| Shortest notice | 3 days (Codex models, 2023) | 60 days | same day (an experimental Maps tool) | \n| Stated minimum | 6 months GA, 3 months variants; previews may get far less | 60 days | 2 weeks for previews; no period stated for stable models | \n\nEach dot is one notice. The medians sit far apart: six months at OpenAI, two at Anthropic and four weeks at Google, where all but two notices fall inside the first two months.\n\n## OpenAI: six months, with three exceptions\n\nOpenAI publishes the most detailed policy of the three. Unless safety or compliance require a faster timeline, its minimum is six months for a generally available model and three for a specialized variant such as a chat or Codex model. Preview models, it says, **“may be retired with much shorter notice, such as 2 weeks.”**\n\nThe record mostly tracks the policy. Ten of the 15 notices covering generally available models landed between 179 and 184 days, and three ran between ten months and a year, including the GPT-4-32k models in 2024. Specialized variants got about three months, 87 to 94 days. Previews, which the policy says may get two weeks, got a median of about five months.\n\nThree exceptions stand out. In March 2023 OpenAI retired its original Codex coding models three days after announcing it. In April 2026 two older dated snapshots of its audio and realtime mini models got 92 days, inside a batch that was otherwise previews and variants. And on September 11, 2026 it gave gpt-5.4-cyber 20 days. The page does not say which tier that model belongs to; as a specialized variant its minimum would be three months, and the policy allows a faster timeline for safety or compliance.\n\nSome OpenAI snapshots also stay available for years. The gpt-4-0613 snapshot, announced on June 13, 2023, is scheduled to shut down on October 23, 2026, more than three years later.\n\n## Anthropic: sixty days, and a date you can see from launch\n\nAnthropic promises [**“at least 60 days’ notice before model retirement for publicly released models.”**](https://platform.claude.com/docs/en/about-claude/model-deprecations) For a while it gave more: Claude 2 and Claude 3 Opus each got about six months in 2025. Since the end of 2025 it has given the minimum, give or take two days, on all four notices.\n\nWhat Anthropic adds is a date you can see coming. Its status table lists a not-sooner-than retirement date for every active model, and for each of the ten models we checked against Anthropic’s launch posts, from Sonnet 4.5 to Opus 5.5, that date is the first anniversary of the launch. Claude Opus 4.1 was retired on its first birthday, August 5, 2026. Claude Sonnet 4 and Opus 4 lasted a year and about three weeks.\n\nSo the planning rule for Claude is simple: assume about a year from launch, expect the notice around month ten, and budget two months to move. Claude models on Amazon Bedrock and Google Cloud follow those platforms’ own schedules, which can differ.\n\n## Google: weeks, even for stable models\n\nGoogle’s deprecations table warns that its dates are [**“the earliest possible dates on which a model might be retired.”**](https://ai.google.dev/gemini-api/docs/deprecations) Its models page promises at least two weeks of notice for previews and gives no notice period for stable models.\n\nThe measured record is weeks for almost everything. Previews got a median of 28 days, from 3 to 50. The stable-model notices ran from 15 to 103 days: Gemini 2.0 Flash got 103, Imagen 4 got 63, text-embedding-004 got 42, and Google’s Veo 2 and Veo 3 models, all of them generally available, got 15, with Veo previews as the recommended replacement in the Gemini API.\n\nTwo preview shutdowns came faster than the two weeks the models page mentions. A preview image-generation model got 3 days in November 2025, and gemini-3-pro-preview got 11 days in February 2026, a little over three months after its launch, after which its ID was pointed at its successor.\n\nSeven more retirements never had a dated announcement in the release notes, so they sit outside the medians. The generally available Gemini 1.5 models appear there only on September 29, 2025, already shut down. Imagen 3 appears the same way on November 10, 2025, and Gemini 1.0 Pro on February 18, 2025 as no longer supported. The rest are the end of model tuning in May 2025, an experimental 2.5 Pro model, two Live API previews, and Gemini 1.0 Pro Vision, which was marked deprecated with no date and removed a month later.\n\nHow long a model lives varies inside every lab. A few well-known examples:\n\n| Model | Lab | Launched | Switched off | Lived | \n|---|---|---|---|---|\n| gpt-4-0613 | OpenAI | Jun 13, 2023 | Oct 23, 2026 (scheduled) | about 3 years, 4 months | \n| claude-sonnet-4 | Anthropic | May 22, 2025 | Jun 15, 2026 | a year and about 3 weeks | \n| claude-opus-4-1 | Anthropic | Aug 5, 2025 | Aug 5, 2026 | exactly 1 year | \n| gemini-2.0-flash |  | Feb 5, 2025 | Jun 1, 2026 | about 16 months | \n| veo-3.0-generate-001 |  | Sep 9, 2025 | Jun 30, 2026 | under 10 months | \n| gemini-3-pro-preview |  | Nov 18, 2025 | Mar 9, 2026 (ID redirected) | 111 days | \n\n## When it is a product, it is a rebuild\n\nOpenAI gives its products long notice, and they are still the most expensive notices to receive. The Assistants API got a full year, with its users pointed to the Responses and Conversations APIs. On June 3, 2026 it announced the end of Agent Builder, the Evals platform and reusable prompt objects, all off by November 30. The Videos API and its Sora 2 models got six months and no replacement.\n\nGoogle’s product changes came faster. A legacy response schema in its Interactions API was removed 33 days after notice, and URL context on Gemini 2.0 Flash ended a week after it was announced.\n\nA model retirement usually means changing one string and rerunning your tests. A product retirement means moving everything you built on it, and for Sora there was no named place to move to. The higher up the stack a lab’s product sits, from hosted assistants to agent builders to eval dashboards, the more of your code a notice reaches.\n\n## Access can close before the shutdown\n\nThree changes in the record stop new users before anything is switched off.\n\n- **OpenAI, May 7, 2026.** Organizations that had not previously run fine-tuning could no longer start jobs from the day of the announcement. From July 2 the same applied to organizations without recent fine-tuned inference, and existing customers lose job creation on January 6, 2027.\n- **OpenAI, June 2024.** Eleven days after announcing the retirement of the GPT-4-32k and vision-preview models, OpenAI limited them to existing users.\n- **Google, September 18, 2026.** Access to the Gemini 2.5 models was limited to users who had actively used them before, with a note saying the 2.5 family is still served and has not been deprecated.\n\nIf you start a new project on an older model, or your product opens new API projects for customers, that door can close with little or no notice.\n\n## What to check before you build on a model\n\n1. **Read the label first.** Previews and experiments account for most of the short notices in this ledger. If paying users depend on a feature, put it on a generally available model.\n2. **Write the earliest retirement date down on day one.** Anthropic lists one for every active model, and Google lists one for only a few. For OpenAI, count on the policy floor instead: six months from a notice for a generally available model.\n3. **Budget the move by lab.** Plan on roughly six months to migrate off an OpenAI model, two off a Claude model, and two to six weeks off a Gemini model.\n4. **Make the model a setting, with a test set behind it.** Keep the model name in config and twenty real prompts with known-good outputs, so a migration is an afternoon of rerunning and comparing.\n5. **Keep hosted products at arm’s length.** Agent builders, hosted assistants, prompt objects and eval dashboards are where a shutdown turns into a rebuild. Keep your prompts, data and logic in your own code.\n6. **Make sure the notice reaches you.** The labs email accounts with active usage, so route that address to an inbox you read, and follow each lab’s deprecations page or release notes.\n7. **For the part that must never break, keep a fallback nobody can switch off.** An open-weight model on your own hardware has no deprecation schedule. Our[Ollama primer](https://okaneland.com/primer/how-to-use-ollama/) covers running one.\n\nOne email, when there's something worth sending\n\n## Get the research in your inbox.\n\nNo fixed schedule, no filler. You get an email when we've tested something, run the numbers, or found a tool worth your time.\n\nFree. Double opt-in, unsubscribe in one click.\n\nGot numbers of your own? [Get the research →](https://okaneland.com/subscribe/)\n\n## Sources & how we researched this\n\n1. OpenAI API deprecations page, read 2026-09-27: deprecation entries from June 2022 to September 2026, each with an announcement date and a shutdown date and most with a recommended replacement (38 notices under our counting rule), plus the stated minimum notice periods: at least 6 months for generally available models, at least 3 months for specialized variants, and possibly much shorter for previews. developers.openai.com/api/docs/deprecations\n2. Anthropic model deprecations page, read 2026-09-27: 9 notices from September 2024 to June 2026, the commitment to at least 60 days of notice for publicly released models, and the model status table with its not-sooner-than retirement dates. platform.claude.com/docs/en/about-claude/model-deprecations\n3. Anthropic launch posts, used for launch dates: Claude 4 (2025-05-22), Opus 4.1 (2025-08-05), Sonnet 4.5 (2025-09-29), Haiku 4.5 (2025-10-15), Opus 4.5 (2025-11-24), Opus 4.6 (2026-02-05), Sonnet 4.6 (2026-02-17), Opus 4.7 (2026-04-16), Opus 4.8 (2026-05-28), Sonnet 5 (2026-06-30), Opus 5 (2026-07-24) and Opus 5.5 (2026-09-22). anthropic.com/news\n4. Google Gemini API release notes, read 2026-09-27: 24 dated shutdown notices from August 2025 to September 2026, plus seven retirements recorded there only after the fact and the Gemini 2.5 access limit of 2026-09-18. ai.google.dev/gemini-api/docs/changelog\n5. Google Gemini API deprecations table, read 2026-09-27: release and shutdown dates, used to confirm shutdown dates where the table lists the model and to date model lifetimes; the page notes that listed dates are the earliest possible retirement dates. ai.google.dev/gemini-api/docs/deprecations\n6. Google Gemini API models page, read 2026-09-27: the model stages, including the at-least-two-weeks commitment for preview models; no notice period is given for stable models. ai.google.dev/gemini-api/docs/models\n7. Okane Land shutdown ledger, built 2026-09-27: one row per announcement date, shutdown date and kind of thing retired, across the records above; notice counted in calendar days; each row linked to its source. Published with this piece as shutdown-ledger.csv.", "url": "https://wpnews.pro/news/openai-gives-six-months-anthropic-gives-two-google-gives-weeks", "canonical_source": "https://okaneland.com/study/ai-model-shutdown-ledger/", "published_at": "2026-10-09 00:00:00+00:00", "updated_at": "2026-10-09 23:28:34.734382+00:00", "lang": "en", "topics": ["ai-products", "ai-tools", "artificial-intelligence", "generative-ai", "ai-infrastructure"], "entities": ["OpenAI", "Anthropic", "Google", "Sora API", "Assistants API", "Agent Builder", "Videos API", "Gemini 1.5"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/openai-gives-six-months-anthropic-gives-two-google-gives-weeks", "markdown": "https://wpnews.pro/news/openai-gives-six-months-anthropic-gives-two-google-gives-weeks.md", "text": "https://wpnews.pro/news/openai-gives-six-months-anthropic-gives-two-google-gives-weeks.txt", "jsonld": "https://wpnews.pro/news/openai-gives-six-months-anthropic-gives-two-google-gives-weeks.jsonld"}}