Data · Lifecycle

AI model deprecation tracker

The shutdown dates that break builds, the replacements providers recommend, and the notice period you actually get — across 12 providers, checked on 20 November 2026.

P
Plugsky Data desk
Baseline 20 November 2026 · reviewed 20 November 2026

Short answer. Two dates matter more than any other in the next quarter. OpenAI shuts down the GPT-5 family, o3 and o3-pro on 11 December 2026, and retires GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini on 14 January 2027. Anthropic retires Claude Sonnet 4.5 on 30 November 2026. If a pinned model ID appears anywhere in your code, check it against the tables below.

Baseline: 20 November 2026. Notice and shutdown dates are read from provider deprecation pages on that date. A model missing from this page is not necessarily safe; it means we have not verified a notice.

How to read the tracker

StatusMeaningWhat to do
ActiveNo retirement announcedNothing, but keep the ID in configuration
DeprecatedA shutdown date is publishedMigrate before the date; requests still work
RetiredRequests now failRemove the ID; check error handling
Partner-managedThe host sets its own datesCheck Bedrock, Vertex or Azure separately

That distinction causes incidents: a deprecated model still answers, so a smoke test passes and nobody migrates, while a retired one returns an error a naive client may retry forever.

Upcoming shutdowns

ShutdownProviderModel or featureNoticeReplacement
2026-11-30Anthropicclaude-sonnet-4-5-202509292026-09-30claude-sonnet-5-5
2026-12-01OpenAIGPT Image 1.5, 1-mini, chatgpt-image-latest2026-06-02GPT Image 2.6 Sunburst or Flare
2026-12-11OpenAIgpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro2026-06-11gpt-6.1-sol, gpt-6.1-terra, gpt-6-luna
2026-12-11OpenAIo3, o3-pro2026-06-11gpt-6.1-sol with pro reasoning
2027-01-06OpenAItts-1, tts-1-hd, gpt-4o-mini-tts2026-10-01gpt-realtime-2.1-mini
2027-01-14OpenAIgpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini2026-10-14gpt-6-luna
2027-01-20OpenAIRealtime and audio family, including gpt-realtime and gpt-audio2026-07-20gpt-realtime-2.1, gpt-audio-1.5
2027-02-10OpenAIgpt-image-22026-11-10GPT Image 2.6 Sunburst or Flare
2027-02-26OpenAIwhisper-1, GPT-4o transcribe family2026-08-26gpt-live-transcribe, gpt-transcribe
2027-04-01OpenAIgpt-5.3-codex, gpt-5.4-nano, gpt-5.12026-10-01gpt-6-sol or gpt-6-luna

OpenAI also lists a 30 November 2026 shutdown for /v1/prompts, the Evals dashboard, and Agent Builder. Those are endpoint and feature retirements, not model retirements, and they break builds in the same way.

Recent retirements

RetiredProviderModelNoticeReplacement
2026-10-23OpenAIgpt-4, gpt-4-turbo, o1, o3-mini, o4-mini, gpt-4o-2024-05-13, gpt-3.5-turbo2026-04-22GPT-5.6 Sol, Terra or Luna
2026-09-28OpenAIgpt-3.5-turbo-instruct, babbage-002, davinci-002, gpt-3.5-turbo-11062025-09-26gpt-5.6-terra
2026-09-24OpenAIVideos API and the Sora 2 family2026-03-24No one-to-one replacement
2026-08-26OpenAIAssistants API2025-08-26Responses and Conversations APIs
2026-08-05Anthropicclaude-opus-4-1-202508052026-06-05claude-opus-4-8
2026-06-15Anthropicclaude-opus-4-20250514, claude-sonnet-4-202505142026-04-14claude-opus-4-8, claude-sonnet-5-5
2026-04-20Anthropicclaude-3-haiku-202403072026-02-19claude-haiku-4-5-20251001
2026-03-26OpenAIgpt-4-0314, gpt-4-0125-preview2025-09-26GPT-5 or GPT-4.1
2026-02-19Anthropicclaude-3-5-haiku, claude-3-7-sonnet2025-12-19, 2025-10-28claude-haiku-4-5, claude-sonnet-5-5
2026-01-05Anthropicclaude-3-opus-202402292025-06-30claude-opus-4-8
2025-10-28AnthropicClaude 3.5 Sonnet (both snapshots)2025-08-13claude-sonnet-5-5
2025-07-21Anthropicclaude-2.0, claude-2.1, claude-3-sonnet2025-01-21claude-opus-4-8, claude-sonnet-5-5
2025-06-06OpenAIgpt-4-32k and both snapshots2024-06-06gpt-4o
2024-11-06AnthropicClaude 1.x and Instant 1.x2024-09-04claude-haiku-4-5
2024-01-04OpenAItext-davinci-003, text-curie-001 and the base GPT-3 models2023-07-06gpt-3.5-turbo-instruct

Notice periods by provider

ProviderPublished noticeWhere it is statedVerified
OpenAI6 months for GA models, 3 months for specialised and image models, as little as 2 weeks for previewsDeprecations pageYes
AnthropicAt least 60 days for publicly released modelsModel deprecations pageYes
Google (Vertex AI)Per-model; the deprecations page lists each model’s shutdown dateVertex AI deprecationsPer model
Microsoft (Azure OpenAI)Per-model lifecycle page; retirement dates are set per deployment typeAzure OpenAI model lifecyclePer model
Amazon (Bedrock)Per-model; Bedrock sets its own schedule, which can differ from the model developer’sBedrock model lifecyclePer model
Meta (Llama weights)No retirement schedule for downloaded weights; hosted versions follow the hostLicence and model cardn/a
MistralPer-model; check the model overview before pinningDocsPer model
CoherePer-model; deprecations are announced in the changelogDocs changelogPer model
AI21Per-model; legacy Jurassic models were retired on a published scheduleDocsPer model
DeepSeekPer-model; the API docs list aliases and retirement noticesAPI docsPer model
Alibaba (Model Studio)Per-model; snapshots and aliases have separate lifecyclesModel Studio docsPer model
xAIPer-model; aliases can move without a version bumpDocsPer model

Two lessons. The notice you get depends on the tier: a preview can vanish in two weeks. And partner platforms are not bound by the developer’s schedule, so a model retired on one API may live longer on another.

Alerts: query the catalogue, do not scrape the blog

OpenAI exposes a shutdown_date field on the models endpoint. Anthropic publishes dates in documentation and by email to active deployments. A robust check reads the catalogue, compares it with the IDs in your configuration, and fails a build when a pinned model is deprecated.

import os, json, urllib.request

PINNED = {"gpt-6-luna", "claude-sonnet-5-5"}
req = urllib.request.Request(
    "https://api.plugsky.com/v1/models",
    headers={"Authorization": f"Bearer {os.environ['PLUGSKY_API_KEY']}"},
)
models = json.load(urllib.request.urlopen(req))["data"]
live = {m["id"] for m in models}
missing = PINNED - live
if missing:
    raise SystemExit(f"pinned model no longer served: {sorted(missing)}")

Run it on a schedule and on every deploy. It turns a 3 a.m. 404 into a failed build with a name in it.

Methodology

  • Source order. Provider deprecation page first, provider changelog second, partner platform page third. A date without one of those is not published here.
  • Dates. The shutdown date is the last day the model accepts requests. The notice date is the announcement date, not the first rumour.
  • Scope. Text, image, audio and realtime models are included when they have a public ID. Fine-tuned variants follow their base model and are not listed separately.
  • Coverage. Twelve providers. Coverage is uneven: OpenAI and Anthropic publish schedules, several others publish only per-model notes.

This tracker is not a substitute for your own inventory. It tells you what the provider announced; only your configuration knows what you call.

Sources

Links checked 20 November 2026. We cite the primary page for each figure. Where a provider publishes a price behind a login, we say so and do not guess.

FAQ

Which AI models are being retired in December 2026?

OpenAI retires the GPT-5 family, o3 and o3-pro on 11 December 2026, and the GPT Image 1.5 family on 1 December 2026. Anthropic retires Claude Sonnet 4.5 on 30 November 2026. All three dates are published by the providers and were checked on 20 November 2026.

What is the difference between deprecated and retired?

A deprecated model still accepts requests but has a published shutdown date. A retired model no longer accepts requests and returns an error. Deprecated is the warning; retired is the outage.

How much notice do providers give?

OpenAI promises at least six months for generally available models, three months for specialised and image models, and as little as two weeks for previews. Anthropic promises at least 60 days for publicly released models. Partner platforms set their own schedules.

Does a retirement on one platform affect the same model on Bedrock or Vertex?

Not necessarily. Amazon Bedrock and Google Vertex AI set their own retirement dates, which can differ from the developer’s API. If you call a model through a partner platform, track that platform’s lifecycle page.

How can I detect a deprecation automatically?

Read the provider’s models endpoint and compare the returned IDs with the IDs pinned in your configuration. OpenAI also returns an optional shutdown_date field. Fail the build when a pinned ID disappears or gains a shutdown date.

Why is a model I use missing from the tracker?

Because we have not verified a public notice for it. Absence is not a guarantee of availability. Send the provider page and we will add the row with its source date.

Use the data, then test it on your own workload

Plugsky publishes its plan caps and model limits. Start free with no card, or compare the flat plans against per-token billing.

Start Free → See plans