OpenAI is un-shipping fine-tuning, check your calendar

Fine-tuning used to be a product. As of this month it is a countdown.

The deprecations page lays it out in three gates. May 7, 2026: any org that never ran a fine-tune is locked out of starting one. July 2, 2026, two weeks ago: if your org has not run inference on a fine-tuned model in the past 60 days, you can no longer create new training jobs either. January 6, 2027: nobody creates new jobs, active customer or not. Inference on models you already trained keeps working until the underlying base model retires, which, given the pace of the deprecation thread, is not a date I would build a roadmap on.

the loophole nobody will admit to using

Read gate two again. The test is inference in the last 60 days. If your nightly classifier retrains weekly but serves constantly, you are fine. If you trained one glorious model in March and let it idle, you are already out. A scheduled heartbeat request against your fine-tune is a dumb, effective way to keep the door open while you plan an exit. I am not proud of this advice. It works.

the actual exits

Three, in order of how often they are the right call. One: a bigger prompt on a current model with caching, which is where most 2023-era fine-tunes should have been all along (the decision guide from last year holds up; the vendor just answered its title question for you). Two: distill onto open weights you host, which trades an API bill for an ops bill, so price the ops honestly. Three: a provider that still sells tuning as a product and says so out loud.

The lesson is not "OpenAI bad". The lesson is that any capability you rent can be un-launched, and the ones with per-org usage gates get un-launched on a schedule you do not control. Audit what you rent.