OpenAI opens up GPT-4o fine-tuning

Developers have spent months cramming massive prompt templates and few-shot examples into every single API call just to get frontier models to follow strict formatting rules.

OpenAI just launched fine-tuning for GPT-4o, giving all developers on paid usage tiers the ability to customize its flagship model with domain-specific datasets.

To jumpstart adoption, OpenAI is offering 1M free training tokens per day for every organization through September 23. If you want to customize GPT-4o mini, you get 2M free training tokens daily over the same window.

Better performance with dozens of examples

According to OpenAI, developers can see meaningful performance jumps with as few as a couple dozen training examples. The customization helps lock down specific tones, output structures, and niche domain instructions.

Training GPT-4o costs $25 per million tokens. Once deployed, inference runs at $3.75 per million input tokens and $15 per million output tokens.

On the data privacy side, fine-tuned models stay under user control. OpenAI confirmed that enterprise data, including training inputs and generated outputs, will not be shared or used to train other foundational models.

Benchmarks are already tumbling

Early partner results show how dramatic the jump can be when pairing frontier weights with targeted data.

Cosine built Genie, an AI software engineering assistant powered by a fine-tuned GPT-4o model trained on human engineering workflows. Genie scored 43.8% on the SWE-bench Verified benchmark and hit 30.08% on SWE-bench Full, shattering its prior 19.27% score for the largest leap ever on that benchmark.

Another partner, Distyl, used fine-tuned GPT-4o to capture the top spot on the BIRD-SQL text-to-SQL benchmark, logging a 71.83% execution accuracy.

Why it matters

We think this launch fundamentally changes the unit economics for developers building vertical AI agents. Instead of bloating every request with hefty system prompts to enforce formatting, baking behaviors directly into model weights drastically cuts runtime latency and token costs.

If you run code generation, SQL pipelines, or enterprise workflows with strict output schemas, this is worth testing immediately, especially with the daily free training credits reducing experimentation risk to zero.

Competitive pressure is mounting as well. Startups have been leaning heavily on competitors like Claude 3.5 Sonnet for coding tasks, but allowing companies to fine-tune GPT-4o gives OpenAI a strong moat against generic frontier APIs.

However, fine-tuning is not magic. It excels at teaching style and execution structure, not teaching brand new world knowledge. Teams that rush in with noisy datasets risk degrading base model reasoning while burning through training budgets.