OpenAI launches GPT-4o fine-tuning, then schedules its shutdown

OpenAI finally gave developers what they’ve been asking for: fine-tuning for GPT-4o. The announcement on August 20, 2024 came with a sweetener — 1 million free training tokens per day for every organization through September 23. At $25 per million training tokens and $3.75/$15 for input/output inference, the pricing isn’t outrageous, though it’s not pocket change either. GPT-4o mini fine-tuning got the same treatment, with 2 million free daily tokens on offer.

The real headline, though, is buried in a May 2026 update at the top of the post. OpenAI is winding down the entire fine-tuning platform. New users are already locked out. Existing customers can still spin up training jobs for a few more months, and fine-tuned models will keep running until their base models get deprecated. But the clock is ticking. If you built something critical on a custom GPT-4o, start planning your exit.

Before the shutdown news, the case studies looked genuinely impressive. Cosine’s Genie, an AI coding assistant, hit 43.8% on SWE-bench Verified using a fine-tuned GPT-4o trained on real software engineer workflows — a massive leap from the previous 19.27% state-of-the-art. Distyl grabbed first place on the BIRD-SQL benchmark with 71.83% execution accuracy. These aren’t marginal gains. They’re the kind of numbers that justify the fine-tuning effort.

OpenAI emphasized the usual safeguards: your data stays yours, automated safety evaluations run continuously, and usage monitoring is in place. But with the platform already in its final months, the question isn’t whether fine-tuning works. It’s whether the long-term bet pays off when the rug gets pulled on the infrastructure itself. For developers who invested in custom models, the timeline in that doc link is now the most important reading on the page.