The shortest artificial intelligence (AI) model retirement notice on the market is 2 weeks, and it breaks your audit trail. The model version your audit log names has an end date, and the vendor picks it. The proof you built for a regulator has a shelf life you did not set.

The big picture:

Pinning a model version is how a governed AI agent stays reviewable. The record names the exact model, the prompts, and the tool calls. A reviewer can re-run the decision months later to see why the system made a choice.

That works until the vendor pulls the version. My read is that the retirement notice is the weak link in enterprise governance.

Most teams treat a model retirement as a buying event. It is actually an evidence event. Firms get a fixed number of days to re-test an agent. A vendor sets that clock, and the buyer cannot negotiate it.

By the numbers

  • 2 weeks — Shortest published notice: OpenAI can retire preview models with just 2 weeks of warning. This sits below six months for generally available (GA) models and three months for specialized variants [1].
  • 60 days — Anthropic floor: Anthropic provides at least 60 days’ notice before a public model retires, for customers with active deployments. Partner platforms set their own timelines [2].
  • 12 months — Google Cloud floor: Google lists models as available for at least 12 months after release. It prints a firm retirement date that will not move earlier [3].
  • 15 days — Bedrock inactivity rule: Amazon Bedrock models in a Legacy window block new users. Existing customers may lose access after 15 days of inactivity, before the end of life (EOL) date [4].

What I’d watch:

The teams closest to this issue are not arguing about notice lengths. They are moving the proof away from the model itself.

  • Hashing the version: A record that just says “Model X, Version 2026-09” is worth little. The view at this desk is that the flight recorder keeps the weights and version, the temperature, the prompt hashes and the tool arguments. A retirement kills the served endpoint, but a hash lets a reviewer verify the exact state [1].
  • Pricing the clock: A team pinned to a 12-month model runs a slower review cycle than a team on a 45-day Legacy clock [3][4]. The number I watch is the shortest notice window across the stack, and the channel can change it [2].
  • Watching the feed: OpenAI lists a shutdown date for each model in its application programming interface (API) [1]. A governance team can read a retirement there instead of waiting on email.
  • Treating swaps as re-tests: Anthropic tells teams to test the replacement model before the old one goes dark [2]. Operators treat this test as the exact same work as the first sign-off.

The catch

None of these numbers is an ironclad promise. OpenAI applies its notice floors unless safety or compliance requires a faster cutoff [1]. Google commits to its printed dates but stays quiet on safety exceptions [3]. Amazon Bedrock sets its own dates entirely, meaning one model can hold several different expiry dates at once [4].

The part I keep circling: pinning buys proof, but it costs freshness. A pinned agent misses the vendor’s latest safety patches and quality fixes. Every pin is just a future migration you booked for yourself.

The alternative carries a quieter cost. Always jumping to the newest version destroys the audit trail. You lose the exact model that made last quarter’s decision.

There is a channel trap, too. Buying the same model through a cloud reseller can change both the clock and the state [2][4]. A risk register that names the model maker but not the platform that served it logs a fact that binds nobody.

At a glance

  • The Big Shift: Vendors now retire model versions on published schedules that run from 2 weeks to 12 months. A firm that pins a version to preserve an audit trail holds an asset with an expiry date set by the vendor [1][2][3][4].
  • Why It Matters: A record stops being replayable the moment its model version vanishes. The real risk is finding out at audit time that a key automated decision cannot be rebuilt.
  • What I’d Watch: I am watching how operators split the proof from the live model endpoint.
  • Version hashing: Recording the exact weights and prompt hashes so the audit log outlives the model.
  • Channel-aware registers: Logging the specific cloud platform that served the model, since that platform owns the retirement date.
  • Re-test budgets: Sizing test capacity to match the shortest notice window in the tech stack, not the longest.
  • The Fine Print: Notice floors are vendor claims with safety carve-outs, not service guarantees. Pinning freezes an agent on an older model. The control that saves your audit trail also delays your security fixes.

Related reading

Sources

[1] OpenAI, “Deprecations” — model deprecation notice periods and the per-model shutdown date. https://platform.openai.com/docs/deprecations [2] Anthropic, “Model deprecations” — 60-day notice floor, lifecycle states, partner-platform schedules. https://docs.claude.com/en/docs/about-claude/model-deprecations [3] Google Cloud, “Model versions” (Vertex AI) — availability of at least 12 months after release, per-model retirement dates, no earlier dates. https://cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versions [4] Amazon Web Services, “Model lifecycle” (Amazon Bedrock) — Active, Legacy and end-of-life states, 6-month and 45-day Legacy periods, 15-day inactivity rule. https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html