OpenAI enters Jev’s turf with a decision API for AI agents

Reuters

OpenAI has launched an application programming interface (API) that helps AI agents make quick, low-cost choices from a list of set options. This puts the company in competition with TypeSafe's Jev, one of the first AI models designed for these kinds of tasks.

The Decisions API became available to everyone on October 22. It is designed for tasks where an AI application does not need a detailed answer from a model and only needs to choose from a few options.

Whenever an AI agent sorts information into categories, picks a tool or decides what to do next, it sends a separate request to a model. If a large model handles every small decision, the time and cost can add up across the whole process.

What are decision models?

Jev helped make this type of AI model, known as a decision model or System One model, more popular. The name comes from psychology, where it describes fast, automatic thinking. Unlike general-purpose models, which create detailed answers, decision models focus on making quick choices. They can also estimate how confident they are in each decision.

How is OpenAI's approach different?

Using a general-purpose model such as OpenAI's GPT-6 Luna to decide whether a customer's question should be sent to a higher-level support team, or which team should handle a request, uses more computing power than the task needs.

With the Decisions API, developers set a question and list the possible answers. The API returns a choice in a structured format, along with a score or probability. The application then uses this result to decide what to do next.

Instead of building a separate decision model, OpenAI has adapted GPT-6 Luna to make these choices more quickly and cheaply. The company says the API can make decisions up to 10 times faster than using Luna through its Responses API.

Nikunj Handa, OpenAI's product lead for APIs, told TechCrunch that decision times had dropped from one to three seconds to between 100 and 180 milliseconds.

OpenAI charges $0.10 for every million input tokens, which are units of text that a model processes, and $0.01 for every million cached input tokens. It does not charge for output tokens.

Where is it being used?

Possible uses include sorting financial transactions into categories, sending customer support requests to the right teams and making language-learning apps more interactive. OpenAI already uses the API in its Dots voice assistant to respond faster. Its personal finance team also uses it to categorise transactions.

The API also works with OpenAI's GPT-Live voice system. For example, a customer support app could use it during a live conversation to decide whether to ask another question, send the request to a different team or pass the problem to a higher-level support team.

Speak, an AI-powered language-learning platform and an early customer, told TechCrunch that it had reduced the time between actions in its app from 1.4 seconds to 0.13 seconds.

Can OpenAI compete with Jev?

Speed and cost are not the only things that matter. A decision model must also make the right choice reliably. Tests cited by TechCrunch show that models from Liquid and Base Lab performed well in decision-model evaluations. OpenAI's model ranked lower on two measures, while Jev remained competitive.

For developers, the real test will be how well these models perform on their own tasks, especially when the information they receive is unclear.