Yesterday, 23 October 2026, OpenAI switched off GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, o1, o3-mini, o4-mini and GPT-4.1 nano in its API, together with the fine-tuned versions built on them. On 30 November Anthropic retires Claude Sonnet 4.5, and on 11 December OpenAI retires GPT-5 and o3. Every model a company builds on now has an expiry date, and it comes sooner than most project plans assume. We compared the official deprecation pages of Anthropic, OpenAI, Mistral and Google to show how long models live, how much warning you get, and how to plan migrations instead of fighting fires.
Deprecated, legacy, retired: what the terms mean
The providers use almost the same vocabulary. A legacy model no longer receives updates. A deprecated model still works, but it is no longer recommended and has a retirement date. A retired (OpenAI: "shut down", Google: "shutdown") model is gone: requests to it return an error. The phase between deprecation and retirement is your migration window, and its length is what the policies differ on.
- Anthropic notifies customers with active deployments by email and in the documentation, with at least 60 days' notice for publicly released models. Active models carry a "not sooner than" retirement date, roughly one year after launch (Claude Opus 4.6: 5 February 2027; Claude Opus 5.5: 22 September 2027).
- OpenAI gives at least 6 months for generally available models, at least 3 months for specialised variants (chat, Codex, deep research) and about 2 weeks for preview models. The
GET /v1/modelsendpoint now returns ashutdown_datefield. - Mistral lists deprecation and retirement dates per model version on its models page but states no minimum notice in the table.
- Google lists the earliest possible shutdown dates for Gemini models; preview models typically last a few months.
- Manus is an agent product, not a model API. It publishes no model lifecycle policy: agent versions (Manus 1.5 in October 2025, 1.6 in December 2025) replace each other in the product. The risk there is not failing requests but changed behaviour in workflows you have already built.
How long models live
Anthropic's dated model IDs make lifespans easy to measure. The Claude 3 generation lived 16 to 25 months from snapshot to retirement; Claude Haiku 3 lasted longest, at over two years. Since the Claude 4 generation, the pattern is a year: Claude Sonnet 3.7 and Opus 4.1 lived exactly 12 months, Sonnet 4 and Opus 4 13 months, Sonnet 4.5 will reach 14. With new frontier models every few months, providers retire old ones to free up compute.
Source: Anthropic model deprecations page (platform.claude.com), October 2026. Start = date in the model ID (snapshot), end = retirement date on the Claude API. Claude Sonnet 4.5 retirement scheduled for 30 November 2026. Amazon Bedrock and Google-managed Vertex AI set their own dates.
Anthropic is also open about the cost of this. Its deprecation page lists the downsides (users who value a specific model must migrate, researchers lose access, safety and model-welfare risks) and points to commitments published in November 2025: the weights of all publicly released models are preserved for at least the lifetime of the company, and each model is interviewed about its preferences before retirement.
How much warning you get
The decisive number for planning is the notice period. We measured it for every deprecation notice with dates: 10 at Anthropic, 24 model and endpoint notices at OpenAI in 2025-2026, and 46 Mistral model versions.
Sources: deprecation pages of Anthropic (10 notices, September 2024 to September 2026), OpenAI (24 model and endpoint notices, April 2025 to October 2026) and Mistral (46 retired model versions with both dates). Our calculation. OpenAI's 20-day minimum is the gpt-5.4-cyber model; Mistral's 0 days is Codestral Mamba (deprecated and retired on 6 June 2025).
Anthropic's notices have converged on the policy minimum. Early retirements (Claude 2, Claude Opus 3) had six months; since December 2025 every notice has given 60 to 62 days. For a company, two months is enough to swap a model ID, but tight for re-testing prompts, re-running evaluations and getting approval in a regulated process.
Source: Anthropic model deprecations page, deprecation history (notification date to retirement date). Highlighted: notices at or near the 60-day minimum.
The pace is rising
The number of retirements grows with the number of releases. OpenAI published 3 deprecation notices in 2024, 13 in 2025 and 15 so far in 2026, and a single notice can cover a dozen models: the April 2026 notice behind yesterday's shutdown listed GPT-3.5 Turbo, GPT-4, o1, o3-mini, o4-mini, gpt-image-1 and their fine-tuned variants. Mistral deprecated 10 model versions in 2024, 15 in 2025 and 21 in 2026, including Mistral Medium 3, Mistral Small 3.2, Magistral and Devstral. OpenAI is also retiring whole products: the Assistants API shut down on 26 August 2026, and Agent Builder, the Evals platform and reusable prompts follow on 30 November.
Sources: OpenAI deprecations page (notices by announcement date), Mistral models page (model versions by deprecation date), Anthropic model deprecations page (model IDs by notification date, including Claude Mythos Preview). 2026 to 24 October. Units differ per provider: one OpenAI notice can cover many models.
Shutdown calendar to early 2027
| Date | Provider | What retires | Recommended replacement |
|---|---|---|---|
| 30 Nov 2026 | Anthropic | Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) | Claude Sonnet 5.5 |
| 30 Nov 2026 | OpenAI | Agent Builder, Evals platform, reusable prompts | Agents SDK; prompts in your code |
| 1 Dec 2026 | OpenAI | gpt-image-1-mini, gpt-image-1.5 | gpt-image-2.5 |
| 11 Dec 2026 | OpenAI | GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3, o3-pro | GPT-5.6 family |
| 6 Jan 2027 | OpenAI | New fine-tuning jobs for remaining customers; tts-1 snapshots | None for fine-tuning |
| 14 Jan 2027 | OpenAI | GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini (and their fine-tunes) | gpt-6-luna |
| 20 Jan 2027 | OpenAI | gpt-realtime, gpt-audio and older realtime models | gpt-realtime-2.1 |
| 26 Feb 2027 | OpenAI | whisper-1, gpt-4o-transcribe | gpt-transcribe |
| Open | Anthropic | Claude Haiku 4.5 (not sooner than 15 Oct 2026), Claude Opus 4.5 (not sooner than 24 Nov 2026) | Notice of at least 60 days |
Model IDs are not the only thing that changes. Anthropic's API version 2026-10-21 removed the temperature, top_p and top_k parameters: requests that still send them get a 400 error. Code that ran unchanged for two years can break on a parameter, not a model.
A migration playbook for companies
| Step | What to do | Why |
|---|---|---|
| Inventory | List every model ID in code, configs, no-code tools and vendor products; export usage by API key (Claude Console) or read shutdown_date from OpenAI's models endpoint | You cannot migrate what you do not know you use |
| Pin, then plan | Pin dated snapshots in production, aliases only in prototypes; put every "not sooner than" date in the team calendar | Stable behaviour, no surprise changes |
| Evaluation set | Keep 50 to 200 real cases with expected outputs and run them on every candidate model (see our guide to LLM-as-a-judge evaluation) | A two-month window is enough only if testing is automated |
| One place for the model | Route calls through one gateway or config, so a migration is one change (see model routing) | Model IDs scattered across services multiply the work |
| Fine-tunes and embeddings | Plan re-training or re-embedding with the base model's retirement | Fine-tuned models die with their base; new embeddings mean re-indexing |
| Cloud platforms | Check Bedrock, Vertex AI and Azure dates separately | Partner platforms set their own schedules |
| Contracts | Ask vendors of AI products which model they use and how they announce changes (see AI vendor selection) | Your supplier's migration becomes your behaviour change |
| Budget | Reserve two migration windows a year per critical AI system | At a 12-month lifespan, one migration a year per model is the norm |
What this means for companies in Croatia and the DACH region
Many companies in the region built their first AI projects in 2024 and 2025 on GPT-4o, GPT-4.1 or Claude 3.5 and 3.7, often through an agency that has since moved on. Those models are retired or will be by January 2027. A short audit answers three questions: which models run where, what the next retirement date is, and whether there is a test set to prove that the replacement works. In regulated sectors (finance, insurance, health) and for high-risk systems under the EU AI Act, a model change can also mean updating documentation, so the 60-day window is shorter than it looks. Treat models like any other dependency with a support end date: plan updates, test them and record them.
Frequently asked questions
How much notice do AI providers give before retiring a model?
Anthropic promises at least 60 days for publicly released models, and its last six notices were 60 to 62 days. OpenAI promises at least 6 months for generally available models, 3 months for specialised variants and about 2 weeks for previews; its median notice in 2025-2026 was about 180 days. Mistral publishes no minimum: its median is about 80 days, and 17 of 46 retired model versions had a month or less.
How long does a Claude model stay available?
In the Anthropic API, Claude 3 models lived 16 to 25 months from their snapshot date to retirement. Claude 4 models (Sonnet 4, Opus 4, Opus 4.1, Sonnet 4.5) lived 12 to 14 months. Active models carry a "not sooner than" retirement date, roughly one year after launch. Amazon Bedrock and Google-managed Vertex set their own dates.
What happens to fine-tuned models when the base model is retired?
They are retired with it. OpenAI states that fine-tuned versions follow the shutdown schedule of their base models, and it retired fine-tuned GPT-3.5, GPT-4 and o4-mini models on 23 October 2026. OpenAI also stops new fine-tuning jobs for remaining customers on 6 January 2027.
Should we pin model versions or use aliases?
Pin dated snapshots in production so behaviour does not change under you, and use aliases only in prototypes. Pinning means you must migrate before the retirement date, so pair it with an inventory of model IDs, an evaluation set and a calendar of shutdown dates.
Is a model retirement coming for your AI system?
We audit which models your applications and vendors use, set up evaluation sets and migrate to the current Claude, GPT or Mistral models without surprises. As a Claude Certified Architect from Zagreb, we work with companies in Croatia and the DACH region.
Talk to an AI consultant