Week 30, 2026: the July 23 shutdown is not the one you think - and we got it wrong last week

AI Stack Radar · published 2026-07-20 · covers 2026-07-16 → 2026-07-20 · 6 action items · 1 correction

The rule of this digest: if it doesn't break your build, change your bill, or carry a deadline - it's not here. About the radar →

CORRECTIONOpenAI's July 23 wave is the current-generation tier - including gpt-5-codex

Last week we told you the July 23 shutdown covers legacy gpt-3.5-turbo, gpt-4, o1 and o3-mini snapshots. We re-checked the source this week. That was wrong, and it was wrong in the direction that hurts. The 2026-04-22 announcement splits into two waves, and the wave label does not track how old a model is:

Retiring Thursday, July 23 (18 IDs): gpt-5-chat-latest, gpt-5-codex, gpt-5.1-chat-latest, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini, gpt-5.2-codex, computer-use-preview and computer-use-preview-2025-03-11, o3-deep-research and o3-deep-research-2025-06-26, o4-mini-deep-research and o4-mini-deep-research-2025-06-26, gpt-4o-search-preview-2025-03-11, gpt-4o-mini-search-preview-2025-03-11, gpt-4o-mini-tts-2025-03-20, gpt-audio-mini-2025-10-06, gpt-realtime-mini-2025-10-06.

Retiring October 23 (the actually-old tier): gpt-3.5-turbo and variants, gpt-4/gpt-4-0613/gpt-4-turbo and their completions variants, gpt-4-1106-preview, gpt-4.1-nano, gpt-4o-2024-05-13, o1, o1-pro, o3-mini, o4-mini including ft-o4-mini-2025-04-16, and gpt-image-1.

Do: if you read last week's version and concluded "we don't run gpt-3.5, not our problem", re-check today. Codex-family coding agents, computer-use flows, deep-research pipelines and the March 2025 search-preview models are all in Thursday's wave. Repoint codex and chat-latest aliases to gpt-5.5, computer-use and the mini tier to gpt-5.4-mini, deep-research to gpt-5.5-pro, and the dated audio/realtime/tts snapshots to their 2025-12-15 equivalents or gpt-audio-1.5.

Source: OpenAI deprecations (verified 2026-07-20). Our OpenAI calendar and the W29 digest have been corrected.

DEPRECATIONAnthropic: legacy Workbench and three prompt-tools endpoints retire Aug 17

Announced July 17. The legacy Workbench at platform.claude.com/workbench is being sunset, and the experimental prompt tools APIs go with it: /v1/experimental/generate_prompt, /v1/experimental/improve_prompt and /v1/experimental/templatize_prompt. After removal, requests to these endpoints return an error. Saved prompts, variables and evals are not supported in the updated Workbench.

Do: two separate jobs. Export saved prompts, variables and evals from the banner or under Organizational Settings before August 17 - they do not carry over. Then grep your code for /v1/experimental/: if prompt generation or templatizing sits inside a build step or an agent loop, this fails in CI before you notice it in production.

Source: Anthropic API release notes

DEPRECATIONAnthropic: claude-mythos-preview retires tomorrow, July 21 - but only on the Claude API

The first-party retirement lands July 21 with a migration path to claude-mythos-5. The wrinkle for multi-cloud setups: Azure Foundry lists claude-mythos-preview (gated research preview) with a retirement date of 2027-04-02. Same model string, two surfaces, nine months apart.

Do: repoint Claude API traffic to claude-mythos-5 today. If you route across providers, do not let the Azure date lull you - the failure will look like "it works in one region and 404s in another".

Sources: Anthropic model deprecations · Azure Foundry retirement schedule

DEPRECATIONAzure Foundry: mistral-document-ai-2505 retires today, July 20

The 2505 preview of Mistral Document AI reaches its retirement date today, with mistral-document-ai-2512 and mistral-ocr-4-0 listed as replacements. Our own correction: this calendar listed August 17 for this model until today. It was July 20.

Do: if you have an OCR or document-extraction path pinned to the 2505 preview, move it to 2512 or mistral-ocr-4-0 now and re-test extraction quality - OCR model swaps change field-level output more than they change headline accuracy.

Source: Azure Foundry retirement schedule (verified 2026-07-20)

DEPRECATIONAWS Bedrock: four Claude 3.x Sonnet variants hit EOL July 30

Still open from last week, now 10 days out: Claude 3 Sonnet, Claude 3.5 Sonnet (20240620), Claude 3.5 Sonnet v2 (20241022) and Claude 3.7 Sonnet reach end of life on Bedrock. Claude 3 Haiku follows on September 10, and Claude Sonnet 4 on October 14.

Do: check your Bedrock model IDs against the lifecycle page this week and move remaining traffic to Sonnet 4.6 or Sonnet 5. If you deferred this after last week's digest, July 30 is now inside a normal sprint.

Source: Bedrock model lifecycle

WATCHopenai-python 2.46.0 removes beta annotation compatibility aliases

Released July 17. The changelog lists "remove beta annotation compatibility aliases" under bug fixes, not under breaking changes - so we are flagging it as a watch item rather than calling it breaking. If your code still imports beta-era annotation aliases, a routine dependency bump can stop compiling.

Do: pin your current version, upgrade in a branch, and run your type checker before merging. Do not let this one ride in on an automated dependency PR.

Source: openai-python releases

Checked and quiet this week

Nothing digest-worthy from Google Gemini (latest changelog entry: July 6), Mistral (latest: June 30), or the Anthropic Python SDK beyond feature additions in 0.117.0 (dreaming support, MCP Tunnels - launches, not obligations). Anthropic's deprecation history added no new entries in this window.

One thing we could not verify

OpenAI's public pricing page returned HTTP 403 to our fetch on 2026-07-20, so no OpenAI price claims are made in this edition. We publish what we verified this week, not what we remember from last week. If pricing changed in this window, it will appear in W31 once the page is reachable.

Next 30 days - from the deprecation calendar

Get this weekly - free

One page a week: only what breaks your build or changes your bill. Signup is a $0 Gumroad checkout (email delivery unlocks as the list grows; the digest is always published here).