LLM model deprecation dates for Claude, OpenAI and Gemini

A model deprecation is a provider’s notice that a model ID is no longer recommended and will stop working, usually on a stated date. The retirement, which OpenAI and the Gemini API call the shutdown, is the day requests to that ID start to fail. Each provider lists its dates on its own lifecycle page, and Amazon Bedrock, Google Cloud and Microsoft Foundry keep separate schedules. Of the dates we last checked on 5 November 2026, these are the next big retirements:

• 19 November 2026: gemini-3.6-flash on Vertex AI and the seven o-series models on Azure OpenAI
• 30 November 2026: claude-sonnet-4-5-20250929 on the Claude API and claude-sonnet-4-5 on Google Cloud and Microsoft Foundry
• 11 December 2026: gpt-5-2025-08-07 and o3-2025-04-16, with their gpt-5 and o3 aliases, on OpenAI’s API
• 14 January 2027: gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 on OpenAI’s API

Last checked: 5 November 2026.

These dates back the model-update step of our guide to production-ready AI agents. The tables list every text or reasoning model that a provider has deprecated, retired or shut down with a date after 19 April 2026 or still to come, plus active models whose earliest possible retirement date falls in 2026. Audio, image, video and embedding models are summarised after each provider’s tables. We checked every row on the provider’s own page on 5 November 2026, except one Google Cloud row whose pages Google has removed. That row says so. The same dates are in a JSON file, one record per model ID and platform, licensed CC BY 4.0, for your own alerts.

What each provider means by deprecated, legacy, retired and shut down

The words around an LLM model retirement differ by provider, and so do the notice periods.

Google’s old Vertex AI deprecations page says Vertex AI’s services are now part of Gemini Enterprise Agent Platform and its documentation is no longer updated, so the Vertex AI rows come from the Agent Platform docs.

Each date below is a deadline for a regression run as well as for a configuration change. Our AI agent maintenance retainer runs that comparison before you switch.

Dates for the most-asked models

Before you switch on any of these dates, run your eval suite on the old and new model side by side, as in our LLM regression testing runbook.

Claude Sonnet 4.5 retirement date

claude-sonnet-4-5-20250929 retires from the Claude API on 30 November 2026. Anthropic deprecated it on 30 September and names claude-sonnet-5-5 as the replacement. On Amazon Bedrock it went Legacy on 8 October 2026 and reaches end of life on 8 April 2027. Google Cloud now lists claude-sonnet-4-5 as deprecated from 29 September 2026 and shuts it down on 30 November 2026, the same day as the Claude API.

GPT-5 deprecation date

gpt-5-2025-08-07 and the undated gpt-5 alias shut down on OpenAI’s API on 11 December 2026, with gpt-5.6-sol as the replacement. OpenAI now lists the alias beside the snapshot. Azure OpenAI retires gpt-5 version 2025-08-07 on 9 February 2027.

o3 deprecation date

o3-2025-04-16 and the undated o3 alias shut down on OpenAI’s API on 11 December 2026, with gpt-5.6-sol as the replacement. OpenAI deprecated them on 11 June 2026. o3-pro-2025-06-10 and o3-pro go the same day, replaced by gpt-5.6-sol with reasoning.mode: pro, and o3-deep-research shut down on 23 July 2026. Azure OpenAI retires o3, o3-pro and o3-deep-research earlier than OpenAI does, on 19 November 2026.

GPT-4o deprecation dates

gpt-4o-2024-05-13 shut down on 23 October 2026, and the gpt-4o-2024-08-06 and gpt-4o-2024-11-20 snapshots follow on 14 January 2027 (replacement gpt-6-luna). On Azure OpenAI, version 2024-05-13 retires on 14 January 2027 and the other two on 14 April 2027.

GPT-4.1 deprecation date

gpt-4.1-2025-04-14 and gpt-4.1-mini-2025-04-14 shut down on OpenAI’s API on 14 January 2027. OpenAI deprecated them on 14 October 2026 and names gpt-6-luna as the replacement. gpt-4.1-nano shut down on 23 October 2026. On Azure OpenAI, gpt-4.1, gpt-4.1-mini and gpt-4.1-nano version 2025-04-14 retire on 14 April 2027.

Gemini 2.5 deprecation dates

Vertex AI’s retirement date for gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite was 20 October 2026, and Google hasn’t extended it, though its page still lists the three models in its main table. On the Gemini API they aren’t deprecated, but since 18 September 2026 only existing users can call them.

Gemini 3.6 Flash deprecation date

On Vertex AI, gemini-3.6-flash is a short-term model that retires on 19 November 2026, with gemini-3.8-flash as the replacement. On the Gemini API it has no shutdown date, and requests to the deprecated gemini-3.5-flash are routed to it.

Azure OpenAI model retirement dates

Azure OpenAI, part of Microsoft Foundry, keeps its own schedule, and its dates can fall before or after OpenAI’s. All seven o-series models (o1, o1-pro, o3, o3-pro, o3-mini, o4-mini and o3-deep-research) retire on 19 November 2026. gpt-4o version 2024-05-13 retires on 14 January 2027 and gpt-5 version 2025-08-07 on 9 February 2027. The other two GPT-4o versions, GPT-4o mini and the three GPT-4.1 models follow on 14 April 2027. Microsoft says retirement dates aren’t extendable, and standard deployments move to the replacement at retirement unless versionUpgradeOption says otherwise. Every row is in the cloud platforms section.

Claude model deprecation and retirement dates

Anthropic’s dates apply to the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry, which it calls Anthropic-operated platforms. Amazon Bedrock and the Google-managed Claude models on Google Cloud “set their own retirement schedules” (Anthropic’s model deprecations page and Claude Platform on AWS), so their Claude dates are in the cloud platforms section.

The earliest date for claude-haiku-4-5-20251001 passed on 15 October 2026, and by 5 November neither it nor claude-opus-4-5-20251101 (not sooner than 24 November 2026) had a deprecation notice. Under the 60-day minimum, a notice sent on 5 November would put either retirement on 4 January 2027 at the earliest. Every other active Claude model has a tentative date in 2027, starting with claude-opus-4-6 (not sooner than 5 February 2027) and claude-sonnet-4-6 (17 February 2027).

OpenAI model deprecation and shutdown dates

OpenAI’s next text-model shutdown is on 11 December 2026, for the GPT-5 and o3 snapshots below, and its 11 June notice now lists their undated aliases, such as gpt-5 and o3, beside them. Its newest notice, dated 14 October 2026, sets 14 January 2027 for the other GPT-4o and GPT-4.1 snapshots, with gpt-6-luna as the replacement. Fine-tuned models follow their base model’s schedule. Aliases listed beside a snapshot appear in brackets after it.

Upcoming OpenAI shutdowns

OpenAI models shut down since 19 April 2026

When we checked on 19 October, OpenAI’s page also named ft-gpt-4, ft-gpt-3.5-turbo, ft-gpt-4.1-nano-2025-04-14, ft-babbage-002 and ft-davinci-002 for 23 October 2026, each with a base model as the replacement. It no longer lists them, and says only that fine-tuned versions follow their base model’s schedule.

OpenAI’s GPT-5 and GPT-4o model pages name gpt-5-2025-08-07 and gpt-4o-2024-08-06 as the default snapshots behind the aliases gpt-5 and gpt-4o. The 11 June notice now lists gpt-5 beside its snapshot, so the alias shuts down on 11 December 2026 too. The 14 October notice lists snapshots only, so it doesn’t say what gpt-4o does after 14 January 2027.

Other OpenAI dates on the same page:

• Agent products. Agent Builder, the Evals dashboard and API, and the v1/prompts API are scheduled to shut down on 30 November 2026, and existing evals have been read-only since 31 October 2026. The Assistants API shut down on 26 August 2026.
• Realtime and audio. gpt-realtime, gpt-4o-realtime, gpt-audio, gpt-4o-audio and their mini versions shut down on 20 January 2027 (replacements gpt-realtime-2.1, gpt-realtime-2.1-mini and gpt-audio-1.5).
• Speech and transcription. tts-1, tts-1-hd and both gpt-4o-mini-tts snapshots go on 6 January 2027 (replacement gpt-realtime-2.1-mini). whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize and /v1/audio/translations go on 26 February 2027.
• Images and video. gpt-image-1 shut down on 23 October 2026, and gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest go on 1 December 2026. Sora 2 and the Videos API shut down on 24 September 2026 with no replacement.

Gemini model deprecation dates on the Gemini API and Vertex AI

Google publishes separate dates for the Gemini API and for Vertex AI, and for the 2.5 models they disagree. Vertex AI’s retirement date for gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite was 20 October 2026, and Google hasn’t extended it. On the Gemini API the same models aren’t deprecated, but since 18 September 2026 Google has limited them to users who have used them before, with no shutdown date (release notes, 18 September 2026).

Gemini API

Watch the two routed IDs. Requests to them are “automatically routed to” the newer models, so an agent pinned to gemini-3.5-flash already runs gemini-3.6-flash. The only Gemini API text model with an upcoming date is gemini-3.1-flash-lite, earliest shutdown 7 May 2027, replacement gemini-3.5-flash-lite.

Vertex AI (Gemini Enterprise Agent Platform)

The platforms also split on the 3.x Flash models. gemini-3.6-flash has no shutdown date on the Gemini API but retires on Vertex AI on 19 November 2026, and gemini-3.7-flash, routed on the Gemini API, stays on Vertex AI until 28 January 2027. Vertex AI keeps gemini-3.5-flash until 19 May 2027 or later.

Other Google dates:

• Live API and speech. On the Gemini API, gemini-3.1-flash-live-preview and gemini-2.5-flash-native-audio-preview-12-2025 (replacement gemini-3.8-live) and the speech previews gemini-3.1-flash-tts-preview, gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts have an earliest shutdown of 17 November 2026. Vertex AI retires gemini-live-2.5-flash-native-audio on 13 December 2026.
• Managed agents. The Gemini API’s deep-research-pro-preview-12-2025 agent was scheduled to shut down on 23 October 2026 (replacement deep-research-preview-04-2026).
• Embeddings. Vertex AI retires text-embedding-005, text-embedding-004, text-multilingual-embedding-002 and multimodalembedding@001 on 1 April 2027, and the Gemini API lists gemini-embedding-001 for 14 May 2028 at the earliest.
• Video. The Gemini API’s veo-3.1 preview models and gemini-omni-flash-preview had an earliest shutdown of 22 October 2026 (replacement gemini-omni-1.1-flash).
• Images. On the Gemini API, gemini-2.5-flash-image has an earliest shutdown of 15 March 2027 (replacement gemini-3.1-flash-lite-image), and Google names gemini-nano-banana-2.1 as the replacement for gemini-3.1-flash-image without setting a date. The Imagen 4 models’ earliest shutdown was 17 August 2026.

Amazon Bedrock, Google Cloud and Microsoft Foundry dates

Cloud platforms keep their own schedules, and the gaps run to months.

• Amazon Bedrock uses Active, Legacy and EOL (end of life). For Bedrock usage, “only the dates on this page apply” (Model lifecycle (Legacy)). Models launched before 7 September 2026 stay Legacy for at least 6 months, while newer ones show an earliest EOL date and a 6-month or 45-day Legacy period on the model card (Model lifecycle). In Legacy, new customers can’t adopt the model, existing customers may lose access after 15 days of inactivity, and prices can rise in the later public extended access period. After EOL, requests fail. Dates differ by region, so read modelLifecycle from GetFoundationModel in yours.
• Google Cloud lists shutdowns of the Claude models it manages on its partner model deprecations page, and each Claude model card gives a retirement date, usually an earliest one. Claude Models on Google Cloud (Anthropic managed) is a separate service that Anthropic operates, and Anthropic’s dates apply to it.
• Microsoft Foundry, which includes Azure OpenAI, uses Preview, GA, Legacy, Deprecated (existing customers only) and Retired, when inference returns 410 Gone. GA models get at least 60 days’ notice and previews 30. Most GA models retire 18 months after launch and Anthropic’s after 12, and Microsoft says “Retirement dates aren’t extendable” (lifecycle and support policy). Standard deployments are upgraded to the replacement at retirement unless versionUpgradeOption says otherwise, but provisioned deployments aren’t. In the Models API, Deprecating means deprecated and Deprecated means retired.

The differences most likely to change a plan, from those pages on 5 November 2026:

Anthropic says its dates apply to Microsoft Foundry, yet Microsoft’s model retirement schedule, last updated 23 September 2026, sets fixed dates for two Claude models that Anthropic lists as active, and names claude-sonnet-5-5 as a replacement without listing it. The schedule also predates OpenAI’s notices of 1 and 14 October 2026.

Checklist for 30, 14 and 7 days before a retirement

Start earlier when the notice allows. Anthropic gives at least 60 days and Vertex AI at least 45 for short-term models, but OpenAI’s example notice for a preview is 2 weeks. With less than 30 days, do the 30-day and 14-day items together. The method for each step is in the LLM regression testing runbook.

30 days before

• Find every place the retiring ID is sent: production config, fallbacks, batch and scheduled jobs, LLM-as-a-judge graders, and every cloud platform and region. The usage export by API key and model in Anthropic’s Claude Console helps, as do the shutdown_date field from OpenAI’s GET /v1/models, the modelLifecycle field from Bedrock’s GetFoundationModel, and lifecycleStatus and deprecation.inference from Azure’s Models API.
• Write down the earliest retirement date across those platforms and regions. That is the deadline.
• Pin exact IDs for the current model and each candidate in one configuration value, using dated snapshots, not aliases or latest.
• Choose one or two candidates and record each one’s own retirement date.
• Fix every request the candidate rejects, such as parameters it no longer accepts, before you measure quality.
• Run the full regression suite on the current model and each candidate in the same window, five samples per case, and run the current model against itself once to see the noise floor. If the suite isn’t built from your own failures yet, start with a golden dataset from production failures.

14 days before

• Compare the runs per failure class against thresholds agreed before the run, and read the trace of every lost case.
• Check cost per case and p95 latency against written budgets.
• Decide go, go with conditions or no-go, and record who decided.
• Start a staged rollout with internal users, then a small share of traffic, reading production traces for the classes that lost cases.
• Rehearse one rollback in staging.

7 days before

• Finish the rollout to all traffic with at least 3 working days left before the retirement date.
• Remove the old ID from fallbacks, batch jobs, graders and config, and search the codebase for the string.
• Write the rollback for after the date. A retired ID can’t be the rollback, so name a second pinned model that passed the suite.
• Alert on errors from the old ID in case something still calls it. Retired Claude models fail, Bedrock requests fail after EOL and Azure returns 410 Gone.
• Record the runs, the decision and the dates in the repository.

Our view

These are our recommendations, not the providers’.

Treat the earliest date on any platform in your request path as the deadline, failover included. Claude Sonnet 4.5 leaves the Claude API on 30 November 2026 but stays on Bedrock until 8 April 2027. A Bedrock fallback could then keep serving a model your suite no longer tests, so we’d move the main path and the fallback in the same release.

Check the replacement’s own date before you migrate to it. Anthropic named claude-haiku-4-5-20251001 as the replacement for Claude Haiku 3, and Haiku 4.5’s own earliest date passed on 15 October 2026. On Azure, Microsoft names gpt-5.1 as the replacement for gpt-4o versions 2024-05-13 and 2024-08-06, which retire on 14 January and 14 April 2027. But Azure retires gpt-5.1 on 13 July 2027, and OpenAI’s own API shuts it down on 1 April 2027. We’d run the suite on the newest model the provider offers as well as the named replacement, and choose the one that passes with the latest retirement date.

Don’t let routing or automatic upgrades choose the model. The Gemini API now routes gemini-3.5-flash and gemini-3.7-flash to newer models, and Azure upgrades Standard deployments at retirement. Both change the model behind an unchanged name without a regression run. We’d change the model string ourselves after a passing run, and run the suite on a schedule between releases.

Keep preview models out of customer-facing paths unless you can migrate within 2 weeks. OpenAI gives 2 weeks as an example of the shorter notice previews can get, and Microsoft force-upgrades preview deployments with at least 30 days’ notice.

What this page won’t tell you

• Dates under private agreements. OpenAI says some customers can provision dedicated capacity after a shutdown date, and Bedrock mentions private arrangements with the provider for continued access. Those dates aren’t public.
• Every regional date. Bedrock lists only the regions covered by an announced EOL date, and a model can stay Active in others. Azure upgrades region by region, and Azure Government clouds usually carry one version of a model at a time. Check your region through the platform’s API.
• Behaviour changes inside an unchanged ID or alias. Gemini latest aliases are “hot-swapped with every new release”, with 2 weeks’ email notice only for breaking changes. The routed Gemini IDs and Azure’s automatic upgrades change the model behind a name, and OpenAI’s notices don’t say what gpt-4o does after its default snapshot shuts down. A suite run only when a notice arrives misses these, so run it on a schedule too.
• Whether a date will hold. Vertex AI dates can be extended. OpenAI and Microsoft can retire a model faster for safety, security or compliance reasons. Gemini API dates are earliest possible dates, so a model can run past them.
• Models outside these tables. We list text and reasoning models from Anthropic, OpenAI and Google, plus the cloud differences above. Other families on Bedrock and Foundry, such as Meta’s Llama and Mistral’s models, have their own dates on those platforms’ pages.

If you’d rather we did this

Our maintenance retainer re-runs your regression suite after every model release and retirement notice, reviews production traces monthly and fixes drift, from USD 3,000 a month for one agent.

Sources

All pages were accessed on 5 November 2026, except three Google Cloud model cards, last read on 19 October 2026 and since removed.

• Anthropic, Claude Platform Docs: Model deprecations (undated, latest entry 30 September 2026) and Claude Platform on AWS (undated).
• OpenAI, OpenAI API docs: Deprecations (undated, latest entry 14 October 2026), and the GPT-5 and GPT-4o model pages (undated).
• Google, Gemini API docs: Deprecations (last updated 22 October 2026), Release notes (last updated 30 October 2026) and Models (last updated 28 October 2026).
• Google Cloud, Gemini Enterprise Agent Platform docs: Model versions and lifecycle, Partner model deprecations, Anthropic’s Claude on Google Cloud models and the Claude Sonnet 4.5 model card, all last updated 4 November 2026. The Claude model cards for Sonnet 4, Opus 4 and Opus 4.1 were last updated 15 October 2026 when we read them on 19 October, and returned 404 on 5 November 2026.
• Google Cloud, Generative AI on Vertex AI deprecations (last updated 26 October 2026).
• Amazon Web Services, Amazon Bedrock User Guide: Model lifecycle and Model lifecycle (Legacy) (both undated).
• Microsoft Learn: Microsoft Foundry Models lifecycle and support policy (last updated 24 July 2026) and Model retirement schedule (last updated 23 September 2026).