openai · anthropic · google
The model deprecation
tracker that opens the PR.
Every OpenAI, Anthropic and Google model and API retirement, dated and sourced. Run it on a repo and get every identifier you still call, with the day it stops working.
$npx aidep .no accountthe calendar
What dies next
dates as published by each provider
| identifier | provider | dies | replacement |
|---|---|---|---|
| assistants-api | OpenAI | Aug 26, 2026 | Responses API + Conversations API |
| gemini-robotics-er-1.6-preview | Aug 31, 2026 | gemini-robotics-er-2-preview | |
| videos-api+5 that day | OpenAI | Sep 24, 2026 | none announced |
| babbage-002+3 that day | OpenAI | Sep 28, 2026 | gpt-5.6-terra |
| gemini-2.5-flash-image | Oct 2, 2026 | gemini-3.1-flash-image-preview | |
| ft-babbage-002+15 that day | OpenAI | Oct 23, 2026 | gpt-5.6-terra |
| v1-prompts+2 that day | OpenAI | Nov 30, 2026101 days | none announced |
| chatgpt-image-latest+2 that day | OpenAI | Dec 1, 2026102 days | gpt-image-2 |
already dead
| claude-3-5-sonnet-20241022 | Anthropic | Oct 28, 2025calls fail | claude-sonnet-4-6 |
| gemini-2.0-flash | Jun 1, 2026calls fail | gemini-3.6-flash | |
| chatgpt-4o-latest | OpenAI | Feb 17, 2026calls fail | gpt-5.1-chat-latest |
| claude-opus-4-1-20250805 | Anthropic | Aug 5, 2026calls fail | claude-opus-4-8 |
after install
One PR per retirement date
- 01
Install
Three GitHub permissions, listed in full on the security page. aidep opens one onboarding PR: the complete audit of your repo, every deprecated identifier with file and line. Nothing else happens until you merge it.
onboarding pr12 findings in 7 files 4 dead · 8 dying · nearest 2026-08-26 (5 days) .github/aidep.json added
- 02
Opt in to migration PRs
One PR per retirement date, grouping every file it touches. The safe rewrites are applied; anything that cannot be rewritten safely becomes a manual checklist in the PR body, not a guess. aidep never touches .github/workflows.
your agent briefsites src/chat.ts:41, :88 replace gpt-5.6-sol verified 2026-08-21 trap /v1/prompts dies 2026-11-30
- 03
Merge on evidence
Turn on evals and a model-swap PR ships a pack that replays your prompts against the old model and the new one, in your CI with your keys, and posts the result. That comment is the merge decision.
github-actions · #241gpt-4o-2024-05-13 -> gpt-5.6-sol held on 18/20 prompts held summarize_invoice held classify_ticket drifted extract_line_items
why deterministic matters: openai’s own migration guide points at prompt objects that are themselves deprecated (dead 2026-11-30). aidep inlines configs instead.
the obvious question
Why not just ask my agent?
29 / 174
vendor-named replacements that are themselves already deprecated
Do. It is good at the rewrite. Two things it structurally cannot do: it cannot tell you about a date you never asked about, and it cannot know what shipped after its training cutoff.
Of the 174 registry rows that name a replacement, 29 name a replacement that is itself already deprecated. One in 6. An agent working from last year’s docs lands on a dead target that often, then says “done.”
aidep tells it when, and tells it what is true today. handbook →
what it can touch
Three permissions, and findings only
Metadata read, contents read and write, pull requests read and write. That is the whole list, so aidep can never write .github/workflows; the eval workflow ships as a file you move yourself.
Your tarball is fetched, scanned in memory and discarded. What persists is a path, a line, and the identifier matched there. Never source.
No customer model key, ever. Eval runs happen in your CI with your keys. The one egress is our own Anthropic key, used to draft eval cases, and only when your repo opts in.
Install once. Hear from us when a date gets close.
The register watches your repos and opens the PR. Or leave an email and we will write when the next shutdown date gets close, and nothing else.
free on every repo, migration prs included · $39 an org a month adds the eval run · pricing