Your AI Model Comes With an Expiration Date Now

The View

Look at the dates baked into the model names and do the subtraction. claude-sonnet-4-5-20250929 retires on November 30, 2026. That's fourteen months. gpt-5-2025-08-07 goes on December 11, so sixteen. gpt-4.1-2025-04-14 makes it to January 14, which is twenty-one. The old champion is whisper-1, which came out in early 2023 and lasts until February 2027, if I recall the launch right.

Windows XP got about twelve years of support. A frontier model snapshot gets one to two.

Is that bad? Not exactly. The new models are usually better and often cheaper. OpenAI's suggested replacement for gpt-4o is gpt-6-luna, which lists at 10 cents per million input tokens and 50 cents per million output. gpt-4o was around $2.50 and $10 when I last looked (it has been a while). And Anthropic says plainly that it retires models to make room for new ones, which I'd take over a polite fiction.

So the deadline isn't the problem. The problem is that a model is now a dependency with a shelf life, and most of us still treat it like a library we installed once.

Some of the small sites I run call these APIs. This site makes its card images with an OpenAI model. We moved it off gpt-image-1 on October 11, twelve days before the shutdown, and not because of the shutdown. The newer model just made better pictures. Lucky timing. I'd rather not need luck.

Here's the napkin math. Say you use three models and each lives about eighteen months. On average you're migrating something every six months. Every migration means re-testing prompts, re-checking output formats, and finding the one cron job that still hard-codes the old model name (there's always one). Call it a couple of days per model if you have tests, and a bad week if you don't.

Nobody budgets for that, probably because these dates used to be years apart. They should. Treat it like certificate renewal. Put the retirement date on a calendar the day you pick a model, and keep a small set of real inputs with known-good outputs so a migration takes an afternoon instead of a guessing game.

Anthropic does one thing here I'd like to see everywhere. It promises at least 60 days' notice and says it's preserving retired models' weights, with a hope of making them available again someday. OpenAI's page gives about six months of warning on most of these. Google's "earliest possible" dates are the least comforting of the three.