OpenAI launches Decisions API widely, challenging Jev

OpenAI has made its Decisions API available to all customers for the first time, stepping fully into the market for fast, cheap AI inference that recently turned TypeSafe — a startup a former OpenAI researcher started — into a $7.6 billion company. TypeSafe’s Jev model created the category: a “decision model” that, given a set of options, quickly picks one and attaches a confidence score, functioning as a super-fast classifier rather than a full language model.

Developers quickly found these “System One” models could swap in for LLMs across a range of workflows, running faster and far more cheaply. OpenAI’s product lead for APIs, Nikunj Handa, said his team saw Jev’s traction and immediately spotted huge potential — but rather than train a dedicated decision model from scratch, OpenAI simply reconfigured ChatGPT Luna to run faster and cheaper. “The big unlock with Decisions API has been the speed, shrinking things from one to three seconds to 100, 120, 180 milliseconds,” Handa told TechCrunch — for now, the service still runs on an off-the-shelf version of Luna rather than a purpose-built model.

OpenAI unveiled the Decisions API at its developer day on September 29 and opened it to every customer on Thursday, adding voice and image modalities and working to cut latency by routing inference closer to users around the world. Early trials with customers including Box, Datadog, Ramp and Speak swapped the API in for OpenAI’s own LLMs to boost throughput; Speak, which builds AI language tutors, says the change cut the lag between actions from 1.4 seconds to roughly 0.13 seconds. OpenAI has also used the technology to sharpen the responsiveness of its own Dots voice assistant.

Judging a “best” decision model is murky — optimizing for intelligence, cost, calibration and speed trades off differently by workload, and TypeSafe’s own CEO, Diogo Almeida, tells people to distrust benchmarks and test models on their own tasks, even as TypeSafe published its own rankings alongside a newer Jev release last week that adds image support and “smart-if” logic. On widely cited public benchmarks, Liquid’s new D1 Ultra and Base Lab’s Elephant-1 currently lead; the latest Jev sits seventh and third, while OpenAI’s Decision Model ranks eighth and fourteenth. TypeSafe’s models remain the most-queried decision models on OpenRouter, though OpenAI argues its edge is multimodality — Handa says Luna already handles image classification with strong calibration, and he’s eyeing even lower latency, down toward 60 milliseconds, which could let a model classify every frame of a video stream in real time.

Why it matters

Jev’s own namesake is the economist who first showed that making a resource more efficient tends to make it more ubiquitous, not less used — and that’s the bet underneath this whole category: as decision models get faster and cheaper, expect them to turn up in far more corners of software than today’s use cases suggest.