OpenAI has launched Reinforcement Fine-Tuning (RFT) on its o4-mini reasoning model, introducing a powerful new technique for tailoring foundation models to specialized tasks. Built on principles of reinforcement learning, RFT allows organizations to define custom objectives and reward functions, enabling fine-grained control over how models improve—far beyond what standard supervised fine-tuning offers.
At its core, RFT is designed to help developers push models closer to ideal behavior for real-world applications by teaching them not just what to output, but why that output is preferred in a particular domain.
What is Reinforcement Fine-Tuning?
Reinforcement Fine-Tuning applies reinforcement learning principles to language model fine-tuning. Rather than relying solely on labeled examples, developers provide a task-specific grader—a function that evaluates and scores model outputs based on custom criteria. The model is then trained to optimize against this reward signal, gradually learning to generate responses that align with the desired behavior.
This approach is particularly valuable for nuanced or subjective tasks where ground truth is difficult to define. For instance, you might not have labeled data for “the best way to phrase a medical explanation,” but you can write a program that assesses clarity, correctness, and completeness—and let the model learn accordingly.
Why o4-mini?
OpenAI’s o4-mini is a compact reasoning model released in April 2025, optimized for both text and image inputs. It’s part of OpenAI’s new generation of multitask-capable models and is particularly strong at structured reasoning and chain-of-thought prompts.
By enabling RFT on o4-mini, OpenAI gives developers access to a lightweight yet capable foundation that can be precisely tuned for high-stakes, domain-specific reasoning tasks—while remaining computationally efficient and fast enough for real-time applications.
Applied Use Cases: What Developers Are Building with RFT
Several early adopters have demonstrated the practical potential of RFT on o4-mini:
- Accordance AI built a custom tax analysis model that improved accuracy by 39% over baseline, using a rule-based grader to enforce compliance logic.
- Ambience Healthcare used RFT to enhance medical coding accuracy, boosting ICD-10 assignment performance by 12 points over physician-written labels.
- Harvey, a legal AI startup, fine-tuned a model to extract citations from legal documents with a 20% improvement in F1, matching GPT-4o on performance at reduced latency.
- Runloop trained the model to generate valid Stripe API snippets, achieving a 12% gain using AST validation and syntax-based grading.
- Milo, a scheduling assistant, improved output quality on complex calendar prompts by 25 points.
- SafetyKit boosted content moderation accuracy in production from 86% to 90% F1 by enforcing granular policy compliance through custom grading functions.
These examples underscore RFT’s strength in aligning models with use-case-specific requirements—whether those involve legal reasoning, medical understanding, code synthesis, or policy enforcement.
How to Use RFT on o4-mini
Getting started with Reinforcement Fine-Tuning involves four key components:
- Design a Grading Function: Developers define a Python function that evaluates model outputs. This function returns a score from 0 to 1 and can encode task-specific preferences, such as correctness, format, or tone.
- Prepare a Dataset: A high-quality prompt dataset is essential. OpenAI recommends using diverse and challenging examples that reflect the target task.
- Launch a Training Job: Via OpenAI’s fine-tuning API or dashboard, users can launch RFT runs with adjustable configurations and performance tracking.
- Evaluate and Iterate: Developers monitor reward progression, evaluate checkpoints, and refine grading logic to maximize performance over time.
Comprehensive documentation and examples are available through OpenAI’s RFT guide.
Access and Pricing
RFT is currently available to verified organizations. Training costs are billed at $100/hour for active training time. If a hosted OpenAI model is used to run the grader (e.g., GPT-4o), token usage for those calls is charged separately at standard inference rates.
As an incentive, OpenAI is offering a 50% training cost discount for organizations that agree to share their datasets for research and model improvement purposes.
A Technical Leap for Model Customization
Reinforcement Fine-Tuning represents a shift in how we adapt foundation models to specific needs. Rather than merely replicating labeled outputs, RFT enables models to internalize feedback loops that reflect the goals and constraints of real-world applications. For organizations working on complex workflows where precision and alignment matter, this new capability opens a critical path to reliable and efficient AI deployment.
With RFT now available on the o4-mini reasoning model, OpenAI is equipping developers with tools not just to fine-tune language—but to fine-tune reasoning itself.
Check out the Detailed Documentation here. Also, don’t forget to follow us on Twitter.
Here’s a brief overview of what we’re building at Marktechpost:
Nikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.
