Developers got six months. You got two weeks.

On January 29, 2026, OpenAI announced it was retiring GPT-4o from ChatGPT. It went dark on February 13. Fifteen days.

The same model stayed available in the API.

OpenAI publishes its deprecation policy for developers, and you can read it yourself. Generally available models get at least six months of notice. Specialised variants get at least three months. Preview models, the experimental ones nobody should be depending on, "may be retired with much shorter notice, such as 2 weeks."

Two weeks is what paying ChatGPT subscribers got. The preview-model treatment, for a product they pay for by the month.

I want to be fair here. OpenAI's stated reason was that only about 0.1% of users still chose that model on a given day, and that is almost certainly true. It is also several hundred thousand people, and the ones hit hardest were the ones who had gone deepest. The people who had learned how that particular model answered and written their instructions around it.

Why the gap exists

It is not spite. A developer with an integration has a contract, a support channel, and churn that shows up in somebody's forecast. Your Tuesday morning routine lives in a browser tab. It appears in no dashboard, no renewal risk model, no account review. Nothing about it is visible to the company whose product it depends on.

That is the whole mechanism. You are not being punished. You are being unmeasured.

What it actually costs

If you use AI casually, nothing. The replacement is better than the old one at nearly everything and you will not notice a thing.

If you use it seriously, the cost is real, but it is not the one people name. The output did not get worse. What happened is that months of accumulated knowledge about how to get good work out of one specific model stopped being worth much overnight, and you got a fortnight to find out.

That knowledge felt like skill. Some of it was. A lot of it was familiarity with a set of habits that no longer exist.

The part worth keeping

Anything you have written down that describes the job travels to the next model. Anything that only works because of one model's quirks does not. That is the whole difference, and it decides how much a retirement notice costs you.

Then, when you are weighing any of these tools, ask the question the demos never answer. Not whether it will change. It will. What happens to me when it does?

Stability is a feature. Nobody puts it in a keynote, because there is no way to show it off. You find out whether you had it on the day it is withdrawn, with two weeks' notice, from a product you were paying for.