OpenAI makes its Jev competitor widely available as decision model competition heats up

OpenAI has made its Decisions API generally available, after first announcing it at its September 29 developer day. It's a direct answer to TypeSafe's Jev, the model that created a category of fast, cheap "decision" or System One models: given a set of options, pick one quickly and report how confident you are. That category helped turn TypeSafe, founded by a former OpenAI researcher, into a $7.6 billion company.

Why care? Because a huge share of real software isn't chat. It's routing, classifying, choosing the next action. If those calls drop from seconds to around a hundred milliseconds and get much cheaper, AI moves into places it was too slow or too expensive to go.

We like this category a lot, and we think it's one of the more quietly useful turns in AI this year. OpenAI's product lead told TechCrunch the big unlock was shrinking response times from one to three seconds down to the 100 to 180 millisecond range. Speak, the language-tutoring company, said its time between actions fell from 1.4 seconds to 0.13 seconds. That's the difference between a tutor that feels sluggish and one that feels present.

What's notable is how OpenAI got here. It didn't build a purpose-calibrated decision model. It took ChatGPT Luna and reconfigured it for speed and cost, and it's candid that "plain old Luna" is doing the work for now. That's a fast-follower move, and an effective one, but it shows on the public leaderboards. On the AIM System One Index and the Artificial Analysis Decision Index, OpenAI's model sits eighth and fourteenth, behind Liquid's new D1 Ultra, Base Lab's Elephant-1 and TypeSafe's latest Jev, which ranks seventh and third. TypeSafe's models are also the most queried on OpenRouter.

We'd take those rankings with salt, and so does TypeSafe's own CEO, who urges teams to trial models on their own work. There's no settled definition of a good decision model yet; the trade-off between intelligence, cost, calibration and speed differs by job. That's exactly why we'd urge caution. A model that confidently picks an option in 100 milliseconds is great for routing a support ticket or steering a game character. Plugged into anything with real consequences, calibration matters more than speed, and a confident wrong answer at scale is still wrong, just faster and more often.

Where OpenAI may have the edge is multimodality, with new voice and visual inputs in the public release and its product lead claiming strong calibration on image classification. The stated ambition is around 60 milliseconds, fast enough to imagine a call on every frame of video. That's powerful, and it's also exactly the capability that makes automated surveillance and moderation cheap. Where this leads if it scales is billions of tiny automated judgments nobody reviews.

Our read: a real, useful category, a healthy competitive fight, and a reminder that efficiency drives ubiquity. The model's namesake economist would recognize the pattern. We want the confidence scores to be honest and the high-stakes uses to keep humans in the loop.

Early users of OpenAI's Decisions API included Box, Datadog, Ramp and Speak, and OpenAI used it to improve the responsiveness of its Dots voice assistant. TypeSafe's newest Jev release added image support and "smart-if" statements. A playful benchmark, JevMan, has decision models play Pac-Man; Jev currently tops it.