AI Model Retirement Lint for hardcoded LLM model IDs

Three of the eight model IDs in this llm.config.js already return an error - for the backend engineer who owns the repository the day a vendor's API stops answering.

line 3  [warn]  Shuts down 2026-10-23 on the OpenAI API.                        -> gpt-5.6-sol
line 4  [warn]  Shuts down 2026-10-23 on the OpenAI API.                        -> gpt-5.6-terra
line 5  [warn]  Shuts down 2026-10-23 on the OpenAI API.                        -> gpt-5.6-terra
line 6  [error] RETIRED 2025-10-28 on the Claude API - this request fails now.  -> claude-sonnet-4-6
line 7  [error] RETIRED 2026-04-20 on the Claude API - this request fails now.  -> claude-haiku-4-5-20251001
line 8  [warn]  Shuts down 2026-10-16 on the Gemini API.
line 9  [error] RETIRED 2026-08-26 on the OpenAI API - this request fails now.  -> gpt-transcribe
line 10 [warn]  Shuts down 2026-10-23 on the OpenAI API.                        -> gpt-image-2

That is the whole output, on one file, in the editor. The file that produced it is ordinary:

// llm.config.js - committed by an AI coding assistant
export const MODELS = {
  chat:      "gpt-4-turbo",
  cheap:     "gpt-3.5-turbo-0125",
  reasoning: "o3-mini-2025-01-31",
  claude:    "claude-3-5-sonnet-20241022",
  fast:      "claude-3-haiku-20240307",
  gemini:    "gemini-2.5-flash",
  voice:     "whisper-1",
  images:    "gpt-image-1",
};

Eight lines, eight findings. Three of them fail right now: claude-3-5-sonnet-20241022 was retired on 2025-10-28, claude-3-haiku-20240307 on 2026-04-20, and whisper-1 on 2026-08-26. The other five have dates that have not arrived yet. gemini-2.5-flash shuts down on 2026-10-16. The remaining four are part of the single OpenAI sweep on 2026-10-23 that removes 12 model IDs at once.

A retired model ID is not a syntax error. It is a valid string, it type-checks, every unit test written around it goes green - and it keeps going green until the vendor's API answers 404 instead of a completion, in production, on a date somebody else picked.

Why the assistant that wrote it cannot warn you

This is the part that makes the failure durable. The model that autocompleted claude-3-5-sonnet-20241022 into that config was trained before Anthropic published the retirement notice. Ask it whether the ID is current and it will tell you yes, confidently, because in the world it learned, it was. Its own model ID is frequently on the same list.

So the usual safety nets are all pointed the wrong way. The compiler sees a string. The linter sees a string. The test suite asserts against a mock. The assistant confirms it from the same training year that produced the mistake. Nothing in the loop holds a calendar.

What holds the calendar

AI Model Retirement Lint ships 33 rules, each one carrying a vendor's published date and the vendor's own recommended replacement. It runs over the file you have open and reports line by line, exactly as above.

Three severities separate what is already broken from what is scheduled. error means the call fails today. warn means a vendor has published a shutdown date that has not arrived. info means the vendor has named an earliest possible retirement date - useful when you are deciding what to pin for the next year.

Check the open file free

Checking the file you have open is free and complete: all 33 rules, every date, every replacement, nothing hidden or watermarked. There is also a command that prints the entire dated table so you can check an ID before you type it.

The licence key covers a different job rather than a larger portion of the same one - walking every file in the repository, emitting CI output that fails a build before a dead ID can merge, and rewriting a pinned ID to the vendor's replacement. That is 29 US dollars once.

Every date above was read from the vendors' own deprecation pages on 2026-09-09.