OpenAI is expanding its custom AI training options with Reinforcement Fine-Tuning, or RFT, a method designed to make o1 models more useful in specialized domains. The goal is not just to make a model sound like a dataset, but to help it improve how it works through difficult technical problems.
The approach is aimed at areas where expertise matters and answers often require more than pattern matching. OpenAI points to fields such as law, finance, engineering, and insurance as examples of domains where deep technical knowledge can shape whether an AI system is genuinely useful.
What Reinforcement Fine-Tuning changes
Traditional supervised fine-tuning typically trains a model to reproduce the style, format, and tone of examples it has seen. Reinforcement Fine-Tuning is presented as a different path. According to OpenAI, it can help models develop new ways of “thinking” through problems.
In this setup, the model is given time to work out a solution in an o1-style process. An evaluation system then rates the answer. Successful reasoning patterns are strengthened, while incorrect ones are weakened.
That distinction matters for expert work. In technical domains, the desired behavior is often not just a polished response. The model must connect evidence, apply domain-specific logic, and produce an answer that can stand up to scrutiny.
RFT is therefore positioned as a way to train specialized o1 models with minimal training examples. That could make customization more practical for organizations that have highly specific tasks but do not necessarily have large amounts of training data prepared for every possible case.
Why specialized domains are the target
OpenAI says Reinforcement Fine-Tuning works especially well for specialized fields that require deep technical knowledge. The examples named by OpenAI include law, finance, engineering, and insurance.
These are areas where small differences in reasoning can have large practical consequences. A general answer may be too shallow, while a domain-specific assistant has to follow the structure of the problem and produce output that reflects the field’s expectations.
OpenAI highlights a collaboration with Thomson Reuters as one example. In that work, the compact o1 Mini model was trained to operate as a legal assistant.
The source does not describe the legal assistant’s exact workflow or performance details. What it does make clear is the direction of the work: OpenAI is trying to adapt smaller o1 models for tasks that demand specialized expertise.
A rare disease research example
One of the clearest examples in the source comes from Justin Ree, a bioinformatician at Berkeley Lab. Ree used Reinforcement Fine-Tuning to study rare genetic diseases.
He trained the system using data extracted from hundreds of scientific papers. That data included symptoms and their associated genes, giving the model material tied directly to the task of gene identification.
Ree reports that the RFT-trained o1 Mini outperformed the standard o1 model on this task, even though o1 Mini is smaller and less expensive. He also notes that the model’s ability to explain its predictions makes it particularly useful.
Testing showed that the fine-tuned mini model achieved the highest precision in gene identification. The reported result reached up to 45 percent accuracy at maximum range.
This example illustrates the main promise of RFT as described in the source: a compact model can become more capable on a narrow, technically demanding task when training rewards the reasoning patterns that lead to better answers.
How organizations can access RFT
OpenAI is accepting organizations into its Reinforcement Fine-Tuning Research Program. The program is intended for organizations working on complex tasks that could benefit from AI assistance.
Participants receive access to the RFT API. They can also provide feedback to help improve the API before public release.
OpenAI plans to make RFT more widely available in early 2025. Until then, the research program is the access path described in the source.
For organizations, the practical takeaway is that RFT is being framed as a customization method for difficult, expert-level workflows. It is not described as a general writing-style adjustment. Its emphasis is on teaching models to solve domain problems more effectively through evaluated reasoning.
What to watch next
The source presents Reinforcement Fine-Tuning as part of OpenAI’s broader custom AI training offerings. Its importance depends on whether the method can reliably turn smaller or more efficient models into strong performers for narrow technical tasks.
The early examples point in that direction, especially where the task has clear evaluation signals. A legal assistant trained with Thomson Reuters and a rare disease system trained from scientific papers both show how RFT can be applied to domains with specialized knowledge.
The next step is broader availability. If OpenAI makes RFT more widely available in early 2025 as planned, more organizations will be able to test whether reinforcement-based customization can improve AI assistance in their own complex workflows.