OpenAI launches fine‑tuning for GPT‑4o with free token quota and announces platform sunset
TL;DR
OpenAI has launched fine‑tuning for GPT‑4o, letting developers customize the model with as few as dozens of examples and offering free training token quotas through September 23; the service will be phased out for new users after May 8 2026.
Overview of the Announcement
OpenAI announced that fine‑tuning is now available for its GPT‑4o model. The feature addresses a top developer request by enabling custom versions of GPT‑4o that deliver higher performance and lower inference cost for specific applications. The launch includes a free allocation of 1 million training tokens per day for each organization until September 23, 2024.
How to Get Started
- Fine‑tuning for GPT‑4o is open to all developers on paid usage tiers.
- To create a fine‑tuned model, users navigate to the fine‑tuning dashboard, click Create, and select the base model
gpt-4o-2024-08-06. - Training costs are $25 per million tokens; inference costs are $3.75 per million input tokens and $15 per million output tokens.
- A smaller variant, GPT‑4o mini, is also available. Users select
gpt-4o-mini-2024-07-18and receive 2 million free training tokens per day through September 23. - Detailed usage instructions are provided in OpenAI’s fine‑tuning documentation.
Performance Gains Demonstrated by Early Partners
Cosine’s Genie AI Engineer
- Cosine fine‑tuned GPT‑4o to create Genie, an AI software‑engineering assistant.
- Genie achieved a state‑of‑the‑art (SOTA) score of 43.8 % on the SWE‑bench Verified benchmark, surpassing the previous best of 19.27 %.
- On the full SWE‑bench benchmark, Genie reached 30.08 %, marking the largest improvement ever recorded on that test.
- The fine‑tuned model was trained on real‑world engineering examples and learned to output code patches in a format ready for commit.
Distyl’s BIRD‑SQL Leaderboard Win
- Distyl fine‑tuned GPT‑4o for text‑to‑SQL tasks.
- The fine‑tuned model placed 1st on the BIRD‑SQL benchmark with an execution accuracy of 71.83 %.
- Distyl reported strong performance across query reformulation, intent classification, chain‑of‑thought reasoning, and self‑correction, especially in SQL generation.
Data Privacy and Safety Guarantees
- Fine‑tuned models are fully owned by the developer; OpenAI does not retain or reuse the training data for other models.
- OpenAI runs continuous automated safety evaluations on fine‑tuned models and monitors usage to enforce policy compliance.
Upcoming Platform Sunset
- Update (May 8 2026): OpenAI will wind down the fine‑tuning platform.
- New users can no longer access the platform, but existing users may create training jobs for several months.
- All fine‑tuned models will remain available for inference until their underlying base models are deprecated, as detailed in the official deprecation timeline.
Implications for Developers
- Fine‑tuning GPT‑4o lowers the barrier to building domain‑specific AI applications, requiring only modest datasets.
- The free token allocation accelerates experimentation and early adoption.
- The announced sunset means organizations should plan migration or export strategies before the platform becomes inaccessible.
Next Steps for Interested Developers
- Visit the fine‑tuning dashboard to create a new job.
- Review the pricing and free‑token limits.
- Consult the documentation for best practices on dataset preparation and safety testing.
- Reach out to OpenAI’s custom‑model team for additional customization options.
Authors: Andrew Peng, John Allard, Steven Heidel