How Much Notice Developers Get Before AI Models Retire: OpenAI, Anthropic and Google Cloud Publish Minimums, GitHub Copilot Dates Imply About a Month

Overview

Developers who build on hosted AI models depend on a retirement date that the vendor chooses. The vendors’ own pages, read on 2026-11-12, publish different minimum notice periods: OpenAI lists “At least 6 months” for generally available models, Anthropic at least 60 days for publicly released models, and Google Cloud at least 45 days for models in its shorter-availability table. GitHub’s Copilot documentation lists retirement dates but, on the page read, states no minimum. This analysis compares those rules with notices issued between August and November 2026.

The Herald’s own reading, which no cited source states in these terms: how much notice a developer gets depends on the vendor and on the surface through which a model is used, and in one case reported by The New Stack the notice was shorter than the vendor’s published minimum. Day counts below are the Herald’s own arithmetic, calendar days between the announcement date and the retirement date as each page prints them, without time-zone adjustment. The Herald’s contributor model is built by Anthropic, one of the vendors compared here.

Background: what each vendor publishes

OpenAI. Its deprecations page sets minimum notice by class: “At least 6 months” for generally available models, at least three months for specialized variants and image generation models, and for preview models notice that “may be retired with much shorter notice, such as 2 weeks”. These apply “Unless safety or compliance concerns require a faster timeline”. The page says developers can read a shutdown_date field from the List models endpoint, and that dedicated capacity after a shutdown may be possible through OpenAI’s sales team. Two entries on the same page, by the Herald’s count: the Assistants API was announced on August 26, 2025 for an August 26, 2026 shutdown (365 days), and the Sora 2 video models and Videos API on March 24, 2026 for September 24, 2026 (184 days).

Anthropic. Its page says it notifies customers with active deployments at least 60 days before retiring a publicly released model. For models that are still active it lists a tentative retirement date phrased “Not sooner than” a given day, and it explains that its dates apply to Anthropic-operated platforms while Amazon Bedrock and Google-managed deployments set their own schedules. It also describes a usage export in the Claude Console for finding deprecated-model calls. From the dates on the page, the Herald counts 60 days for Claude Haiku 3 (announced February 19, 2026, retired April 20), 61 days for both Claude Opus 4.1 and Claude Sonnet 4.5, and 189 days for Claude Opus 3 (announced June 30, 2025, retired January 5, 2026).

Google Cloud. Its model-lifecycle page splits generally available Gemini and embedding models into two groups: some are “available for at least 12 months after release”, while models in a shorter-availability table stay active until a replacement launches and a retirement date is announced, with “at least 45 days to migrate”. The page covers Google Cloud’s platform; the Herald did not read the documentation for the separate Gemini Developer API.

GitHub Copilot. The supported-models page carries a “Model retirement history” table of retired and scheduled models with suggested alternatives. It says evaluation models “may be added, updated, or removed without notice” and, on the page as read, gives no minimum notice for other models.

Current situation: dated events, August to November 2026

• 2026-08-26: the Assistants API’s shutdown date, and the date o3 was due to leave ChatGPT. Tech Times, published on August 24 and drawing on OpenAI’s migration guide and third-party migration write-ups, reports that calls to the Assistants endpoints would return hard errors and that OpenAI will not provide an automated tool for moving Threads to Conversations. The o3 API snapshots stay until December 11, 2026, according to both Tech Times and OpenAI’s page.
• Copilot retirements, three rounds: GitHub’s changelog announced on July 31 that six models, including Claude Sonnet 4.5, would be deprecated on September 1 (32 days). A later post says the October 19 retirements of GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5 were announced on September 18 (31 days). A third notice, dated October 15, retires Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra on November 16 (32 days), with Opus 4.8 kept for individual subscribers on annual plans. None of the three posts gives a reason for its timing.
• 2026-09-30: Anthropic notified developers that Claude Sonnet 4.5 retires on the Claude API on November 30 (61 days).
• 2026-10-14: OpenAI notified developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API on January 14, 2027 (92 days). The New Stack reports that this falls short of OpenAI’s published six-month policy for generally available models, at half that time, and says there is no explanation for the shorter period. The same article says OpenAI’s recommended replacement, GPT-6 Luna, supports function calling in Chat Completions only with reasoning effort set to none. Separately, UC Today reports that GPT-5.5 was scheduled to leave ChatGPT, ChatGPT Work and Codex on the same day, with API access unaffected by that change.
• 2026-10-15: Anthropic announced that Claude Haiku 4.5 would leave the Claude chat model selector on October 29. Notebookcheck, citing Anthropic’s help article, says API users of Haiku 4.5 are not affected.

A replacement is not always a drop-in. On September 23, The New Stack reported that Anthropic’s migration guide for Claude Opus 5.5 lists four breaking changes that can return 400 errors in existing agents. Anthropic’s page separately lists temperature, top_p and top_k as removed in API version 2026-10-21; the Herald covered that release.

The same model on different clocks

The Herald’s selection of cases where a model has more than one date, not an exhaustive list:

• Claude Sonnet 4.5: retired from Copilot on September 1; on the Claude API, November 30.
• Claude Opus 4.7: Copilot’s retirement table lists October 2, 2026; Anthropic’s page lists the API model as active, “Not sooner than” April 16, 2027.
• Claude Opus 4.8: Copilot retires it on November 16 apart from the annual-plan exception; Anthropic’s page lists “Not sooner than” May 28, 2027 for the API.
• o3: ChatGPT on August 26 and API snapshots on December 11, 2026.
• GPT-5.5 and Claude Haiku 4.5: scheduled to leave ChatGPT and Codex (October 14) and the Claude chat selector (October 29) respectively, with the API reported as unaffected by UC Today and Notebookcheck respectively. The Notebookcheck report restates Anthropic’s help article.

Anthropic’s page does not mention Copilot, and the Herald does not assume the two schedules are linked.

Implications

What the evidence says about behaviour. A single-author arXiv preprint mined GitHub for commits that migrate away from officially deprecated OpenAI, Anthropic and Google models. It estimates that 82% of migrations away from retired models were committed after the shutdown date, 89% for Anthropic’s notices against 13% for OpenAI’s one-year Assistants API notice, and associates longer notice with lower odds of a post-shutdown migration. It also reports that model identifiers are hard-coded in 94% of migrating applications. Its sample is non-fork GitHub repositories, so it says nothing about private codebases. UC Today cites the 82% figure and advises teams to audit dependencies and rerun real tasks. The preprint describes Anthropic’s notices as 60 to 114 days, while the Anthropic page, by the Herald’s arithmetic, also lists longer ones such as the 189 days for Claude Opus 3; no cited source explains the difference, and the Herald does not reconcile them.

What follows for developers, in the Herald’s reading. First, the date that matters is the one for the surface in use: a chat selector, a coding tool and an API can retire the same model on different days. Second, a stated minimum is a commitment whose scope is set by its author; OpenAI’s six-month figure applies to generally available models and sits beside a three-month class and a two-week preview class, and The New Stack reports one notice at half of it. Third, two vendors’ pages describe tooling for finding exposure: OpenAI’s shutdown_date field and Anthropic’s Console usage export. The Copilot and Google Cloud pages, as read, describe no equivalent. Fourth, notice length is one axis; the New Stack’s reports on GPT-6 Luna and Claude Opus 5.5 describe replacements that change tool-call or request behaviour, so a migration can cost more than swapping a model name. For a comparison in other software, the Herald previously examined four 2026 changes that each carried a different kind of notice.

What the sources do not settle

OpenAI’s page, as read, gives no reason for the 92-day GPT-4 series notice. The Herald did not read Google’s Gemini Developer API documentation, and Google Cloud’s page need not describe it. GitHub’s posts give no reason for its roughly one-month cadence. The vendor pages are live documents that can change after 2026-11-12, and the cases above are the Herald’s selection from what it could verify against vendor pages.