LLM Model Deprecation Tracker

Every announced shutdown date across 3 providers, in one table, with a live countdown and the official replacement for each model. Calling a retired model does not silently fall back — it fails. This is the most common way a working LLM integration breaks in production, and the most avoidable.

53 entries · verified against official deprecation docs 2026-08-18

ModelProviderDeprecatedShuts downTime leftReplacement
claude-3-7-sonnet-20250219
Claude Sonnet 3.7
Anthropic2025-10-282026-02-19—claude-sonnet-4-6
claude-3-5-haiku-20241022
Claude Haiku 3.5
Anthropic2025-12-192026-02-19—claude-haiku-4-5
claude-3-haiku-20240307
Claude Haiku 3
Anthropic2026-02-192026-04-20—claude-haiku-4-5
claude-opus-4-20250514
Claude Opus 4
Anthropic2026-04-142026-06-15—claude-opus-4-8
claude-sonnet-4-20250514
Claude Sonnet 4
Anthropic2026-04-142026-06-15—claude-sonnet-4-6
gemini-3.1-flash-image-preview
Gemini 3.1 Flash Image preview · non-text
Google—2026-06-25—gemini-3.1-flash-image
gemini-3-pro-image-preview
Gemini 3 Pro Image preview · non-text
Google—2026-06-25—gemini-3-pro-image
veo-2.0-generate-001
Veo 2.0 · non-text
Google—2026-06-30—veo-3.1-generate-preview
veo-3.0-generate-001
Veo 3.0 · non-text
Google—2026-06-30—veo-3.1-generate-preview
claude-opus-4-1-20250805
Claude Opus 4.1
Anthropic2026-06-052026-08-05—claude-opus-4-8
imagen-4.0-generate-001
Imagen 4.0 · non-text
Google—2026-08-17—none announced
imagen-4.0-ultra-generate-001
Imagen 4.0 Ultra · non-text
Google—2026-08-17—none announced
imagen-4.0-fast-generate-001
Imagen 4.0 Fast · non-text
Google—2026-08-17—none announced
gemini-robotics-er-1.6-preview
Gemini Robotics ER 1.6 preview · non-text
Google—2026-08-31—gemini-robotics-er-2-preview
sora-2
Sora 2 · non-text
OpenAI2026-03-242026-09-24—none announced
sora-2-pro
Sora 2 Pro · non-text
OpenAI2026-03-242026-09-24—none announced
gpt-3.5-turbo-instruct
GPT-3.5 Turbo Instruct
OpenAI2025-09-262026-09-28—gpt-5.6-terra
gpt-3.5-turbo-1106
GPT-3.5 Turbo 1106
OpenAI2025-09-262026-09-28—gpt-5.6-terra
babbage-002
babbage-002
OpenAI2025-09-262026-09-28—gpt-5.6-terra
davinci-002
davinci-002
OpenAI2025-09-262026-09-28—gpt-5.6-terra
claude-sonnet-4-5-20250929
Claude Sonnet 4.5
Anthropic—2026-09-29 or later—claude-sonnet-5
claude-haiku-4-5-20251001
Claude Haiku 4.5
Anthropic—2026-10-15 or later—claude-haiku-4-5
gpt-3.5-turbo-0125
GPT-3.5 Turbo 0125
OpenAI2026-04-222026-10-23—gpt-5.6-terra
gpt-4-0613
GPT-4 0613
OpenAI2026-04-222026-10-23—gpt-5.6-sol
gpt-4-1106-preview
GPT-4 1106 preview
OpenAI2026-04-222026-10-23—gpt-5.6-sol
gpt-4-turbo
GPT-4 Turbo
OpenAI2026-04-222026-10-23—gpt-5.6-sol
gpt-4.1-nano
GPT-4.1 nano
OpenAI2026-04-222026-10-23—gpt-5.6-luna
gpt-4o-2024-05-13
GPT-4o (2024-05-13)
OpenAI2026-04-222026-10-23—gpt-5.6-sol
o1-2024-12-17
o1
OpenAI2026-04-222026-10-23—gpt-5.6-sol
o1-pro-2025-03-19
o1-pro
OpenAI2026-04-222026-10-23—gpt-5.6-sol
o3-mini-2025-01-31
o3-mini
OpenAI2026-04-222026-10-23—gpt-5.6-sol
o4-mini-2025-04-16
o4-mini
OpenAI2026-04-222026-10-23—gpt-5.6-terra
gpt-image-1
gpt-image-1 · non-text
OpenAI2026-04-222026-10-23—gpt-image-2
claude-opus-4-5-20251101
Claude Opus 4.5
Anthropic—2026-11-24 or later—claude-opus-5
gpt-image-1-mini
gpt-image-1-mini · non-text
OpenAI2026-06-022026-12-01—gpt-image-2
gpt-image-1.5
gpt-image-1.5 · non-text
OpenAI2026-06-022026-12-01—gpt-image-2
gpt-5-2025-08-07
GPT-5 (2025-08-07)
OpenAI2026-06-112026-12-11—gpt-5.6-sol
gpt-5-mini-2025-08-07
GPT-5 mini (2025-08-07)
OpenAI2026-06-112026-12-11—gpt-5.6-terra
gpt-5-nano-2025-08-07
GPT-5 nano (2025-08-07)
OpenAI2026-06-112026-12-11—gpt-5.6-luna
gpt-5-pro-2025-10-06
GPT-5 pro (2025-10-06)
OpenAI2026-06-112026-12-11—gpt-5.6-sol
o3-2025-04-16
o3
OpenAI2026-06-112026-12-11—gpt-5.6-sol
o3-pro-2025-06-10
o3-pro
OpenAI2026-06-112026-12-11—gpt-5.6-sol
gpt-realtime
gpt-realtime · non-text
OpenAI2026-07-202027-01-20—gpt-realtime-2.1
gpt-audio
gpt-audio · non-text
OpenAI2026-07-202027-01-20—gpt-audio-1.5
gpt-4o-realtime
gpt-4o-realtime · non-text
OpenAI2026-07-202027-01-20—gpt-realtime-2.1
gpt-4o-audio
gpt-4o-audio · non-text
OpenAI2026-07-202027-01-20—gpt-audio-1.5
claude-opus-4-6
Claude Opus 4.6
Anthropic—2027-02-05 or later—claude-opus-5
claude-sonnet-4-6
Claude Sonnet 4.6
Anthropic—2027-02-17 or later—claude-sonnet-5
claude-opus-4-7
Claude Opus 4.7
Anthropic—2027-04-16 or later—claude-opus-5
claude-opus-4-8
Claude Opus 4.8
Anthropic—2027-05-28 or later—claude-opus-5
claude-fable-5
Claude Fable 5
Anthropic—2027-06-09 or later—none announced
claude-sonnet-5
Claude Sonnet 5
Anthropic—2027-06-30 or later—none announced
claude-opus-5
Claude Opus 5
Anthropic—2027-07-24 or later—none announced

Countdowns appear once the page loads in your browser, so they are always current rather than frozen at build time.

How each provider handles retirement

Anthropic

Publishes a "not sooner than" floor for every active model and commits to at least 60 days' notice before an actual retirement. Dates apply to the Claude API, Claude Platform on AWS and Microsoft Foundry; Amazon Bedrock and Google Cloud set their own schedules, so a model retired first-party may still run there.

OpenAI

Announces a deprecation date and a shutdown date roughly six months apart, and retires models in large batches — eleven models share the 23 October 2026 date alone. Dated model IDs such as gpt-5-2025-08-07 are retired independently of their undated aliases, so pinning an exact snapshot buys reproducibility at the cost of a fixed expiry.

Google

Announces shutdowns through the Gemini API changelog. Preview models are retired aggressively once a GA equivalent ships — several -preview IDs were discontinued within weeks of their stable counterparts becoming available.

Frequently asked questions

How much notice do LLM providers give before retiring a model?

Anthropic commits to at least 60 days of notice before retiring a publicly released model, and publishes a "not sooner than" date for every active model well in advance. OpenAI has typically given about six months between deprecation announcement and shutdown — for example, models deprecated on 22 April 2026 shut down on 23 October 2026. Google announces shutdown dates in its Gemini API changelog, often with shorter lead times for preview models.

What happens if I keep calling a retired model?

The request fails. There is no silent fallback to a newer model, so an application that hardcodes a retired model ID breaks outright on the shutdown date. This is the single most common way LLM integrations break in production, and it is entirely preventable.

Which models are shutting down soonest?

The table above sorts by shutdown date with a live countdown. As of August 2026 the near-term cluster is: the Google Imagen 4.0 family (17 August 2026), OpenAI Sora 2 and the Videos API (24 September 2026), the legacy GPT-3.5 and completions models (28 September 2026), and a large OpenAI batch including GPT-4 Turbo, GPT-4.1 nano, o1, o3-mini and o4-mini (23 October 2026).

Why do Anthropic dates say "or later"?

Anthropic publishes a "not sooner than" floor rather than a fixed retirement date for active models. The model will not be retired before that date, and customers with active deployments get at least 60 days of notice when a real date is set. Treat these as the earliest possible date to plan against, not a scheduled shutdown.

Do deprecation dates differ between the first-party API and cloud platforms?

Yes. Anthropic states that its published dates apply to the Claude API, Claude Platform on AWS and Microsoft Foundry, while partner-operated platforms — Amazon Bedrock and Google Cloud Vertex AI — set their own retirement schedules. A model retired on the first-party API may remain available on Bedrock or Vertex. Claude Opus 4.1 is a current example.

How do I audit which deprecated models my application still uses?

Anthropic exposes a usage export in the Claude Console (Usage → Export) that breaks down calls by API key and model, which makes it straightforward to find stragglers. More generally, avoid hardcoding dated model IDs across a codebase: route model selection through a single configuration constant so a migration is a one-line change rather than a search-and-replace.

Sources