OpenAI Launches Decisions API After TypeSafe AI’s Jev Debut

OpenAI launched a public beta of its Decisions API on Oct. 6, introducing a structured-output tool for software workflows that need classifications, routing decisions and scores instead of prose responses.

The API accepts text and images and currently runs only on GPT-6 Luna. Its responses can include predicates, fixed-choice classifications or scores that application code can process directly.

The launch follows TypeSafe AI’s early-access release of Jev on Sept. 15. Jev was developed after two years in stealth and reached nearly 13% of paid AI Gateway teams within 24 hours. It reached more than twice as many paid teams in its first day as any previous model launch on the platform.

Jev and Decisions API target repeated, narrowly defined judgments, including sorting support requests, routing documents and determining whether an item should be escalated to a person. Traditional large language models can handle similar tasks, but these systems return constrained outputs designed for automated workflows or human review.

Jev returns typed decisions, probabilities and confidence scores rather than open-ended text. Its input price is $0.042 per million tokens, with no charge for output tokens. Decisions API input pricing is $0.10 per million tokens for GPT-6 Luna, and decision responses do not incur output charges.

Decisions API responses can be delivered up to 10 times faster than responses generated through OpenAI’s general-purpose Responses API. The service is also described as an extension of the existing GPT-6 Luna model rather than a separately trained model.

Jev is part of the “System One Models” approach, which uses a parallel sampler and a training method called Reinforcement Learning for Calibrated Decisions. TypeSafe AI founder Diogo Almeida described the product this way: “Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”

OpenAI API product lead Nikunj Handa said the Decisions API was built on existing model weights. “We haven’t trained like a new model for this,” Handa said. “We’re … building this purely on top of the same Luna weights that we have.”