Anthropic Claude API Deprecation Calendar
Anthropic's risk profile is different from a pure model vendor. The model retirements are gentle and pre-announced with published floors, but the sharp edges are the non-model changes: a memory-listing breaking change under an unchanged header, fast-mode removal, a mid-year pricing step, and parameter validation that now hard-errors. Those are the ones that break a running agent without you touching a model string.
Scheduled changes 2026
| Date | What changes | Do this before the date | Src |
|---|---|---|---|
| Jul 21 | claude-mythos-preview retires | Repoint to claude-mythos-5. Preview aliases were never meant for production - if this one reached prod, add a check that blocks -preview model strings from shipping. Note the surfaces differ: this is the Claude API date, while the Azure Foundry listing of claude-mythos-preview runs to 2027-04-02, so a multi-cloud deployment can keep working on one provider and break on the other. | ✓ |
| Jul 22 | Breaking: old memory-listing semantics change under the existing managed-agents-2026-04-01 header (list/pagination behavior) | This bites even though the header string is unchanged. Update your list/pagination code to the new semantics; current SDKs already send the new behavior, so bump the SDK and re-test any hand-rolled memory listing. Details in the W29 digest. | ✓ |
| Jul 24 | Fast mode (speed:"fast") on Opus 4.7 starts returning errors (4.6 already silently degrades) | Move fast-mode calls to Opus 4.8, or drop the speed parameter. Grep for speed in your request builders - a hard error on a hot path is worse than the silent degradation you may already be eating on 4.6. | ✓ |
| Aug 5 | claude-opus-4-1-20250805 retires (Claude API) | Repoint to claude-opus-4-8. Re-run evals: 4.1 to 4.8 is several generations, so prompts tuned for 4.1 quirks may behave differently. | ✓ |
| Aug 17 | Legacy Workbench retires, and the experimental prompt tools APIs go with it: /v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt. After removal these endpoints return an error. Saved prompts, variables and evals are not supported in the updated Workbench (announced 2026-07-17) | Two separate jobs. (1) Export saved prompts, variables and evals from the banner in the Console or under Organizational Settings - the updated Workbench will not carry them over. (2) Grep your code for /v1/experimental/: any prompt-generation or templatize call is a hard error after Aug 17, so move that logic into your own code or a prompt file in the repo. If you wired these endpoints into a build step or an agent loop, this breaks silently in CI first. | ✓ |
| Aug 31 | Price: Sonnet 5 introductory pricing ends ($2/$10 per Mtok in/out to $3/$15 from Sep 1) | Inputs go up 50%, outputs 50%. Re-forecast your monthly Sonnet 5 spend at the new rate, and factor the ~30% tokenizer difference vs. older Claude if you migrated to hit the intro price. | ✓ |
Retirement floors ("not sooner than")
Anthropic commits that these models will not retire before the dates below. They are floors, not scheduled shutdowns - useful for planning how long a pin is safe.
| Model | Retirement not before | Planning note |
|---|---|---|
claude-sonnet-4-5-20250929 | 2026-09-29 | Safe to pin through Q3; plan a move by early Q4. |
claude-haiku-4-5-20251001 | 2026-10-15 | Cheapest tier - fine to keep on high-volume paths near-term. |
claude-opus-4-5-20251101 | 2026-11-24 | Holds through most of 2026. |
claude-opus-4-6 | 2027-02-05 | Long runway. |
claude-sonnet-4-6 | 2027-02-17 | Long runway. |
claude-opus-4-7 | 2027-04-16 | Long runway (but fast mode already gone - see Jul 24). |
claude-opus-4-8 | 2027-05-28 | Current migration target - safest pin. |
claude-fable-5 | 2027-06-09 | Longest floor published. |
claude-sonnet-5 | 2027-06-30 | Longest floor; note the Aug 31 price change above. |
Standing gotcha, no shutdown date
Non-default temperature, top_p or top_k now return a 400 on Opus 4.7+, Opus 4.8 and Sonnet 5 (older models silently accepted them). If you migrate a request that sets these, strip them or expect a hard failure - this is a common surprise when moving prompts forward. This calendar is maintenance guidance, not legal advice.