OpenAI opens Decisions API to all, taking on TypeSafe's Jev

What's going on

OpenAI opened its Decisions API to all customers on Thursday, October 22, 2026. It was first shown on September 29 at the company's developer conference. With it, OpenAI is entering a market created by the startup TypeSafe and its Jev model. TypeSafe was founded by a former OpenAI researcher and, according to TechCrunch, is now valued at $7.6 billion.

What are decision models?

Jev pioneered a category of deep learning models that work as a kind of super-classifier: you give them a set of options, they quickly pick one and return a confidence score, meaning how sure they are. They're also called decision or "System One" models (a nod to fast, intuitive thinking). Developers found that in many workflows they can replace large language models (LLMs), and do so faster and far more cheaply.

How OpenAI went about it

Nikunj Handa, product lead for OpenAI's APIs, said colleagues saw Jev and thought there were a million use cases for it. His team quickly built its own version. Rather than developing a purpose-calibrated decision model, it took ChatGPT Luna and reconfigured it to make inference (computing the answer) faster and cheaper. The big gain is speed: from one to three seconds down to 100 to 180 milliseconds. Targeted training could come later; for now it's "just regular Luna."

The public version gained new voice and image capabilities. OpenAI also places compute resources in specific regions to reduce latency worldwide.

First customers

In tests with Box, Datadog, Ramp and Speak, developers swapped OpenAI's LLMs for the API to get higher throughput. Speak CTO Andrew Hsu said the time between actions dropped from 1.4 to 0.13 seconds. OpenAI also used the product to improve the response speed of its voice assistant Dots.

Benchmarks and competition

According to TechCrunch, it's unclear what makes a good decision model, other than that it works for your own use case. Intelligence, cost, calibration and speed have to be weighed differently depending on the task. TypeSafe chief Diogo Almeida distrusts benchmarks, yet TypeSafe also published its own rankings last week for the latest Jev, which adds image support and "smart-if" instructions.

On OpenRouter, TypeSafe is still the most queried for now. More entertaining is Opper AI's JevMan benchmark, where decision models play Pac-Man; the latest Jev leads it. Opper executive Felix Wunderlich attributes that to the constant stream of fast decisions within a few hundred milliseconds.

Outlook

Handa sees OpenAI's edge in multimodality: Luna is very strong at image classification and better calibrated. Customers also want images as output. At around 60 milliseconds, one call per video frame would be conceivable. As costs fall, he expects applications to expand, much as the economist Jev is named after described resources being used more efficiently.

Our take

The move shows that specialized, fast models are a market segment of their own, and big vendors are following. For developers, that means more choice on latency and price. The performance figures come mostly from vendors and customers, and the report doesn't give pricing. On the public indexes, OpenAI so far trails the leaders, and how meaningful such benchmarks are is disputed. If you want to try it, compare on your own workloads.