OpenAI did not build a new model to chase the hottest category in AI infrastructure. It took an existing one, tuned it for speed, and shipped it. That is both the appeal and the weakness of the Decisions API.
OpenAI made the API widely available on Thursday, after announcing it at its September 29 developer day. It targets the market TypeSafe opened with Jev, a startup now valued at $7.6 billion and founded by a former OpenAI researcher. Jev popularized what you could call super classifiers, or System One models. You hand them a set of options, and they pick one and attach a confidence score. Developers found these could replace LLM calls in many workflows at a fraction of the latency and cost.
OpenAI’s version is plain Luna, reconfigured for faster inference. Nikunj Handa, product lead for OpenAI’s APIs, says response times drop from one to three seconds to roughly 100 to 180 milliseconds. He also says there is no special training yet, though that may come. The public release adds voice and visual modalities, and OpenAI is routing inference regionally to cut latency worldwide.
Early customers include Box, Datadog, Ramp, and Speak. Speak’s CTO, Andrew Hsu, says the API cut the time between actions from 1.4 seconds to 0.13 seconds. OpenAI also used it to make its Dots voice assistant more responsive. Those are real numbers, and for anything conversational, that gap decides whether a product feels alive or sluggish.
Now the awkward part. On public benchmarks, OpenAI is not leading. Liquid’s new D1 Ultra and Base Lab’s Elephant-1 tend to top the AIM System One Index and the Artificial Analysis Decision Index. The newest Jev model sits seventh and third on those, while OpenAI lands eighth and fourteenth. Jev also remains the most queried option on OpenRouter, and it tops JevMan, a benchmark where models play PacMan. It is silly, but a nonstop stream of high-stakes choices in a few hundred milliseconds is a fair stress test.
Still, nobody agrees on what makes a good decision model. Speed, cost, intelligence, and calibration trade off differently per workload. TypeSafe CEO Diogo Almeida tells people to test models on their own work, even as TypeSafe published its own rankings last week alongside a Jev update that adds image support and smart-if statements.
OpenAI’s best argument is multimodality. Handa says Luna is strong at image classification and better calibrated there, and customers want images as outputs too. He also floats a 60 millisecond target, which would make a call per video frame plausible.
My read: this matters, mostly because OpenAI’s distribution is enormous. If you already run on its stack, trying the Decisions API costs little. If you need the best pure classifier, run your own bake-off against Jev, D1 Ultra, and Elephant-1 first. The category is real, and the competition is now serious.
