Skip to content
SunsetLint

Anthropic's retirement notices shrank from months to about 60 days

October 9, 2026 · Updated November 9, 2026

If you build on Claude, you may have noticed that retirement announcements feel more urgent than they used to. They are. Anthropic's model deprecation page shows a clear shift: announcements made in 2025 gave about four to six months' notice, and announcements made in 2026 have given about 60 to 62 days.

This post sets out the timeline, explains what the shorter window changes, and suggests how to plan for it. All dates come from Anthropic's model deprecations page, last checked on 6 October 2026. Where we only have the month, we give the month.

The timeline

Between October 2025 and November 2026, Anthropic retired, or scheduled the retirement of, ten model IDs:

  • October 2025: Claude Sonnet 3.5 (two IDs)
  • January 2026: Claude Opus 3
  • February 2026: Claude Sonnet 3.7 and Claude Haiku 3.5
  • April 2026: Claude Haiku 3
  • June 2026: Claude Sonnet 4 and Claude Opus 4
  • August 2026: Claude Opus 4.1
  • 30 November 2026: Claude Sonnet 4.5 (announced 30 September 2026)

That's roughly one retirement event every six to eight weeks across the period. If your team has used Claude for a year or more, you have probably lived through at least two of them.

The policy and the practice

Anthropic's policy commits to at least 60 days' notice before retiring a model, and it emails customers that have active deployments of the affected model. It also offers a usage export, so you can audit which models your organization has been calling.

In 2025, practice was well ahead of the policy minimum: four to six months. In 2026, practice matched the minimum. The Sonnet 4.5 announcement on 30 September for a 30 November retirement is a typical example: 61 days.

We don't know the reasons, and we won't guess. What matters for planning is the number.

What 60 days actually allows

Sixty days sounds like plenty. Here is how it tends to get used up in a small team:

  • Days 1 to 7: the email finds its way to engineering. The notice goes to whoever owns the Anthropic account. If that isn't the engineering lead, it gets forwarded, sometimes more than once.
  • Days 7 to 14: scope. Someone searches the code. In a single service, that's an hour. Across many repositories, background jobs and notebooks, it's a few days, especially if nobody is sure which repositories call the API.
  • Days 14 to 30: choose and test the replacement. Even when the provider recommends a replacement, you need to run your own evaluations. Newer Claude models may also reject parameters you send; for example, Claude 4.7 and later reject temperature, top_p and top_k.
  • Days 30 to 45: roll out. Staged, with monitoring.
  • Days 45 to 60: clean up. Fallbacks, defaults, evaluation scripts, infrastructure variables.

That works, but there's no slack. Lose a week to a holiday period, an incident or a busy release, and you're migrating under pressure. With four to six months, the same work could wait for a quiet sprint. With 60 days, it can't.

How to plan for short windows

1. Make sure the notice reaches engineering on day one

Check who receives provider notices. If your provider lets you add notification contacts, add a shared engineering address. If not, set up a forwarding rule from the account owner's mailbox. This costs five minutes and saves days.

2. Know where your model IDs are before the notice arrives

The slowest step is usually scoping. Do it in advance: keep a list of repositories that call AI APIs, and keep model IDs in one config module per service. Our guide on keeping model IDs in one place covers the patterns.

3. Keep a small evaluation set ready

If you have 20 to 50 representative prompts with expected properties, testing a replacement takes an afternoon instead of a week. More on that in testing a model swap with evals.

4. Treat each retirement as a project with an owner

One person owns the migration from notice to clean-up. Our 60-day runbook is a template you can copy into a ticket.

5. Check platform dates separately

If you call Claude through Amazon Bedrock or Google Vertex AI, those platforms set their own retirement dates, which can differ from Anthropic's. Don't assume the first-party date applies to you. See why your Bedrock Claude may retire on a different date.

Is this only an Anthropic pattern?

No provider is immune to short notices. OpenAI gave GPT-4.5-preview about three months' notice, and o1-preview and o1-mini three to six months, according to OpenAI's deprecations page. OpenAI also gave a full year for the Assistants API, sunset on 26 August 2026. Notice periods vary by provider and by product, which is another reason to read each notice rather than assume.

A quick self-check

Ask these five questions about your own setup. If any answer is "no" or "not sure", that's where a 60-day window will hurt:

  1. Would a retirement email reach an engineer within one working day?
  2. Can you list every repository that calls Anthropic's API, Bedrock or Vertex?
  3. Does each service read its model ID from one place, with no hardcoded fallback?
  4. Could you test a replacement model against real cases in one afternoon?
  5. Is there a CI check that stops retired model IDs from coming back?

Each "yes" takes days off the critical path the next time a notice arrives.

The takeaway

Model retirements are now routine maintenance with a fixed deadline. The deadline is roughly two months for recent Claude retirements. Teams that handle them calmly do three things ahead of time: route notices to engineering, keep model IDs in one place, and keep an evaluation set ready.

For the current deadline, see our note on the Claude Sonnet 4.5 retirement on 30 November 2026.