You can now adjust your own organization version of OPENAI’s O4-MINI thinking with reinforcement learning

Openai announced today An account that focuses on the developer on the social network x Third -party programs outside the company can now reach the refinement of reinforcement (RFT) for the new Linguistic Thinking Form O4-MINIEnably allocate a new version of it based on its unique products for an institution, interior terms, goals, employees, operations, and more.

Basically, this ability allows developers to take the available model for the public and modify it to better fit their needs Openai platform information panel.

Next, they can post it through the Openai (API) applications interface, another part of its developer platform, and connect it to internal computers, databases and applications.

Once it is published, if an employee or leader of the company wants to use it through a custom room dedicated or Openai GPT dedicated To withdraw the knowledge of the private company; Or to answer specific questions about the company’s products and policies; Or create new connections and association in the company’s voice, they can do this more easily with their RFT version of the form.

However, one warning note: research showed this The seized models may be more likely to break the prison and hallucinationsSo continue with caution!

This launch expands the company’s model improvement tools along with the SFT control and more flexible control of the complex tasks of the field.

In addition, Openai announced that the supervision control is now supported by the GPT-4.1 Nano model, which is the most affordable and faster prices of the company.

How to help augmented installation (RFT) organizations and institutions?

RFT creates a new version of Openai’s OPENI thinking form, which is automatically adapted to the user’s goals, or the organization’s/institution goals.

It does this by applying the feedback ring during training, developers in large companies (or even independent developers who work on their own) can start simply, easily and easily through Openai developer platform via the Internet.

Instead of training on a set of questions with the correct fixed answers – which is what the learning subject to traditional supervision does – RFT uses a model to record multiple responses for each mentor.

The training algorithm then adjusts the weight weights so that high -grade outputs become more likely.

This structure allows customers to align models with accurate targets such as the “home style” of the institution for communication and terms, safety rules, realistic accuracy, or compliance with internal policy.

To perform RFT, users need:

1. Determine the function of grades or the use of Openai -based class students.
2. Download a set of data with health verification claims.
3. Create API or micro -dashboard training.
4. Monitor progress, review checkpoints, repetition of data or logic.

RFT currently supports thinking models in Series O only and are available for the O4-MINI model.

Early institutions use cases

On its platform, The most prominent Openai many first customers Those who adopted RFT via various industries:

• Artificial Intelligence Agreement Use RFT to adjust a form of complex tax analysis tasks, improve 39 % in accuracy and outpace all the leading models on tax thinking standards.
• Health care atmosphere RFT was applied to the ICD-10 Code Code, raising the performance of the model by 12 points on the doctors ’lines in the Panel Data set.
• Harvey RFT is used to analyze legal documents, improve F1 grades to extract the quotation by 20 % and match GPT-4O in accuracy while achieving inference faster.
• Runloop Models set to create API Stripe icon scraps, using important class students in sentences and the logic of validation of AST, which achieves 12 % improvement.
• Milo RFT has been applied to scheduling tasks, which enhances the right in highly complex situations by 25 points.
• Safetykit RFT is used to impose moderate policies of micro content and increase the F1 model from 86 % to 90 % in production.
• Chipstackand Reuters ThompsonOther partners also showed gains in performance in generating data, legal comparison tasks, and verification workflow.

These situations are often common characteristics: clear task definitions, organized output formats, and reliable evaluation criteria-all of which are necessary to control the effective reinforcement process.

RFT is now available for verified organizations. Openai provides a 50 % discount for teams that choose to share their training data groups with Openai to help improve future models. Interested developers can start using RFT documents from Openai and Dashboard.

Pricing structure and filling

Unlike controlling or preferred control, which is bills for each symbol, RFT is described based on the time the training is actively spent. especially:

• $ 100 per hour of basic training time (wall time during models deportation, classification, updates, and health verification).
• Time is tightened by the second, almost to two twenty places (so 1.8 hours of training will cost the customer $ 180).
• The fees only apply to the work that modifies the form. Waiting lists, safety examination, and inactivity stages are not carried out.
• If the user uses OpenAi models as a pillbing (for example, GPT-4.1), the bills consumed during the grades are released separately at the standard API rates in Openai. Otherwise, the company can use external models, including open source models, as students in the class.

Below is an example of cost collapse:

scenario	Bitter time	Assign
4 hours of training	4 hours	400 dollars
1.75 hours (consistency)	1.75 hours	175 dollars
Training for two hours + one hour (due to failure)	Two hours	$ 200

The pricing model provides this transparency and effective functional design rewards. To control costs, Openai encourages the difference on:

• Use light or effective class students where possible.
• Avoid excessively validation unless it is necessary.
• Start with smaller data collections or run short to calibrate expectations.
• Monitor training using API or temporary dashboard and stop temporarily as needed.

Openai uses the method of preparing bills called “progressive progress”, which means that users are only bills for the model training steps that have been completed and successfully kept.

Should you invest your organization in issuing a custom version of Openai’s O4-MINI or not?

Enhancing control provides a more expressive and and controlled way to adapt language models for use cases in the real world.

With the support of organized outputs, code -based classroom students, and full control of API, RFT offers a new level of customization in publishing the form. Openai’s startup emphasizes the design of studied tasks and a strong evaluation as keys for success.

Developers interested in exploring this method can access documents and examples via Openai’s dashboard.

For institutions that have clearly specific problems and verified answers, RFT provides a convincing method to align the models with operating objectives or compliance – without building RL infrastructure from scratch.