OpenAI brings reinforcement fine-tuning to o4-mini

Fine-tuning reasoning models has always been a tricky puzzle, but OpenAI is finally opening the gates.

OpenAI Developers announced that Reinforcement Fine-Tuning, or RFT, is officially available today for OpenAI o4-mini. Alongside that rollout, the company also enabled standard supervised fine-tuning for its fastest and cheapest model, GPT-4.1 nano.

Teaching reasoning models to specialize

First previewed last December, RFT uses chain-of-thought reasoning combined with task-specific grading to push model accuracy in dense, complex subjects.

Instead of just memorizing text examples, the model learns to think through a problem and gets scored on how well it solves it. OpenAI pointed to AccordanceAI, which used RFT to build a state-of-the-art model tailored specifically for tax and accounting.

For developers looking for pure speed and ultra-low cost rather than heavy math, GPT-4.1 nano now supports supervised fine-tuning, letting teams adapt OpenAI's lightest model to custom data pipelines.

Why it matters

We think this release gives builders two very distinct playbooks for customization.

On the high end, RFT on o4-mini gives enterprises in law, finance, and scientific research a real way to build reasoning-heavy domain experts without losing reasoning rigor. That is a massive leap over standard prompt engineering for high-stakes workflows.

On the low end, supervised fine-tuning on GPT-4.1 nano is all about raw utility. It allows companies to run fast, hyper-cheap tasks like routing and extraction in-house without paying flagship API prices.

The catch is that task-specific grading for RFT is hard work. If your evaluation setup is flawed, reinforcement learning will only amplify bad habits, meaning the barrier to entry for serious RFT remains steep.