OpenAI’s Decisions API takes on TypeSafe’s Jev in the race for 100-millisecond AI choices

Most real software isn’t chat. It is routing, classifying and choosing the next action. Decision models, sometimes called System One models, do that narrow job: given a set of options, pick one quickly and say how confident you are. TypeSafe, founded by a former OpenAI researcher, created the category with Jev, which skips language generation in favor of calibrated decision probabilities (our coverage of Jev’s debut). The approach helped make TypeSafe a $7.6 billion company and spawned a crowd of alternatives within weeks (20 Jev alternatives).

TechCrunch reports that OpenAI has made its Decisions API generally available. Its product lead said the key gain was cutting response times from one to three seconds down to 100 to 180 milliseconds. Speak, the language-tutoring company, said its time between actions fell from 1.4 seconds to 0.13 seconds, the difference between a tutor that feels sluggish and one that feels present. Early users included Box, Datadog, Ramp and Speak, and OpenAI used the API to make its Dots voice assistant more responsive.

How OpenAI got there is telling. It didn’t build a purpose-calibrated decision model. It took ChatGPT Luna and reconfigured it for speed and cost, and openly says “plain old Luna” is doing the work for now. Luna launched in September as part of OpenAI’s half-price GPT-6 efficiency wave (GPT-6 Sol and Luna). It is an effective fast-follower move, and it shows on the 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. TypeSafe’s own CEO urges teams to trial models on their own work rather than trust rankings, which is the right advice. There is no settled definition of a good decision model yet, and the balance of intelligence, cost, calibration and speed differs by job.

That is why caution matters. A model that confidently picks an option in 100 milliseconds is great for routing a support ticket or steering a game character. In anything with real consequences, calibration matters more than speed. A confident wrong answer at scale is still wrong, only faster and more often.

Multimodality could be OpenAI’s advantage: the public release adds voice and visual inputs, and the product lead claims strong calibration on image classification. The stated goal is around 60 milliseconds, fast enough to imagine a judgment on every frame of video. That is powerful, and it is also what makes automated surveillance and moderation cheap. Scaled up, it means billions of tiny automated judgments that no one reviews. TechCrunch has also noted that a fast, cheap constrained-choice model could serve as the watcher that polices misbehaving agents (our analysis), a genuinely useful application. Confidence scores need to be honest, and high-stakes uses need humans in the loop. More in our OpenAI coverage.