Model Sunset Lint - retired AI model IDs, dated

It reads the file you have open and flags every model ID, endpoint and API parameter that a vendor has already switched off or has published a shutdown date for. Each finding names the vendor, the date itself, and the exact replacement string. 109 rules ship inside, covering OpenAI, Anthropic, Google and Azure.

Engineers who own a shipped LLM feature - the person paged when a call starts returning 404 - and reviewers of code an AI assistant wrote. An assistant cannot know a retirement date set after it was trained, so it confidently emits IDs that stopped answering months ago. This catches them at edit time instead of in production.

An assistant answers from training data, and a retirement calendar changes after training ends - that is exactly the fact it cannot have. Dependabot bumps package versions and never reads a string literal like "gpt-4". Type checkers see a valid string. Nothing in the free toolchain carries a date, so the 404 in production is the alarm.

The free tier checks the open file, or only the lines you select, against all 109 rules. No key, no watermark, no usage cap, and no finding is hidden. The licence covers a different job: sweeping the whole repository in one pass, rewriting the retired IDs in place, exporting the dated inventory, and CI output.

A US software engineer averages $70.92 an hour (ZipRecruiter, August 2026). Auditing one repository by hand means grepping every model string and checking it against four vendors’ deprecation pages, then repeating that each quarter. This is $29 once, one licence key per person or team seat.