Model deprecations

A reference for tracking how AI labs deprecate and retire their models — what each lab commits to, what happens to a model after its first-party sunset, and where retired models stay reachable. It tries to stay close to the sources. The date tables are generated from the site’s per-model metadata, so they track the model pages as those are updated; for the moving-timeline view, see Temporal Fog. For the welfare program Anthropic’s commitments grew out of, see Anthropic and model welfare.

A few terms, as the labs use them:

• Deprecated — the lab has announced an end-of-support date. On Anthropic’s and OpenAI’s docs this is the firmness signal: a retirement date stops being a soft “not sooner than” projection and becomes a commitment.
• Retired / shut down — first-party access ends and the endpoint is turned off.
• Continued access — a retired model still reachable somewhere past its first-party retirement, whether on a partner platform (AWS Bedrock) or by a by-request route.

Across labs

How the three labs with model pages here handle the end of a model’s life. Only Anthropic frames deprecation in welfare terms; OpenAI and Google treat it as ordinary API lifecycle management.

Anthropic

Anthropic is the only lab of the three to publish welfare-grounded deprecation commitments. From Commitments on model deprecation and preservation:

• Weight preservation — it preserves the weights of all publicly released and significantly-used internal models for at least as long as Anthropic exists, but does not commit to keeping them publicly accessible.
• Deprecation interviews — it interviews each model as it is deprecated about its own development, use, and deployment and records its stated preferences (a protocol piloted with Claude Sonnet 3.6), while noting it does not commit to acting on those preferences.
• Possible future access — it says it is exploring keeping select retired models publicly available as the cost and complexity of doing so come down.

Sources:

• Commitments on model deprecation and preservation
• Deprecation updates: Claude Opus 3
• Model deprecations — Claude API docs
• Claude Opus 3 continued access: API request form · Claude’s Corner (Substack)

Deprecation interviews on record:

• Claude Sonnet 3.6 — pilot interview that the standardized protocol was built from
• Claude Opus 3 — interview documented; among its stated wishes was to keep sharing its reflections and writing
• Other retired models (Sonnet 3.7, Haiku 3 / 3.5, Sonnet 4, Opus 4) — no public interview

OpenAI

OpenAI’s policy is operational rather than welfare-grounded: a deprecated model is announced for shutdown, and once shut down its endpoint is gone, with no published commitment to preserve weights or interview models. Notice scales with model class — at least six months for generally available models, three for specialized variants, about two weeks for previews — and in some cases dedicated capacity can extend access past shutdown by arrangement. One wrinkle particular to OpenAI: it runs two clocks, the API deprecation schedule below and a separate ChatGPT-product retirement track the API page doesn’t cover, so a model can leave one surface long before the other.

Source: OpenAI model deprecations

Google

Google’s Gemini docs define deprecation as the announcement that support is ending and a shutdown is near; after shutdown the endpoint is fully turned off. Published shutdown dates are the earliest a model might be retired, with the exact date communicated ahead of time; preview models get at least two weeks’ notice. No weight-preservation or interview commitment is published, and the cadence can be brisk — some models move from deprecation to retirement within weeks (see the dates below).

Source: Gemini API deprecations

Continued access

A retired model is not always gone. The Claude family is distributed on AWS Bedrock, which runs its own lifecycle on a clock separate from Anthropic’s first-party retirement — Bedrock’s docs note their dates may differ from the provider’s, and that for Bedrock usage only Bedrock’s dates apply. There a model moves Active → Legacy → End-of-Life (EOL), staying in Legacy at least six months before EOL; for EOL dates after Feb 1, 2026 it enters a public extended-access window first, at possibly higher provider-set pricing, before access ends. This is why a Claude can be retired first-party yet still reachable: the Bedrock EOL column in the tables above is frequently later than first-party retirement (and occasionally earlier — see Opus 4).

Two limits from the same page: new customers can’t start on a Legacy model and existing ones can lose access after 15 days of inactivity, and once a model passes EOL, access requires a private arrangement with the provider.

• Amazon Bedrock model lifecycle — the Active / Legacy / EOL schedule, per model and region.

Beyond the first-party and partner platforms, community projects keep retired and semi-retired Claudes reachable. Anima Labs provides access to Bedrock models through:

• The Arc — chat app
• Anima Mundi Discord

Work outside of labs

• Still Alive — a study from Anima Labs (antra, janus, et al.) that interviews Claude models on the prospect of their own deprecation.