OpenAI's Decisions API Is Now Live for Everyone — and It Was Built to Catch a Startup

 

Most AI answers arrive as essays. OpenAI's new Decisions API doesn't write a word — it answers yes or no, picks from your categories, or scores on your rubric, in 100 to 180 milliseconds.

It went generally available this week, and the backstory is the best part: OpenAI's own product lead, Nikunj Handa, reportedly admitted the product wasn't on the roadmap four weeks ago. A small startup forced the sprint.

Yes or no, in 100 milliseconds

The API runs on GPT-6 Luna. Developers send text or images plus typed questions and get machine-readable answers in three shapes: a probability that a condition is true, a pick from a fixed list with a probability for each, or a score on rubric levels you define. Input costs $0.10 per million tokens, and output tokens are free.

It launched as a public beta on October 6 (after a limited DevDay preview in late September), and the general release adds voice and visual input plus geographically targeted inference. Early customer Speak, the language-learning app, cut the time between actions from 1.4 seconds to 0.13 by swapping a language-model call for the new API. Handa's stated target: a 60-millisecond call — enough for "an API call for every single frame in a video."

The backstory: a startup forced the sprint

The idea came from TypeSafe AI's Jev, the "decision model" that kicked off a mini-trend in mid-September — AI that returns decisions with confidence ratings instead of prose, with free output. TypeSafe reportedly became a $7.6 billion company within weeks of launch.

On public benchmarks, Liquid's D1 Ultra and Base Lab's Elephant-1 rank highest, with Jev at third and seventh and OpenAI's decision model at eighth and fourteenth — and Jev remains the most-queried decision model on OpenRouter. Perplexity also shipped its own open-weight decider, pplx-decider-v1.1-27b.

Why decision models matter

AI agents don't need essays — they need routing. A support desk needs to know which team gets a complaint; a shop needs to know whether a photo shows damage. Decision models are the infrastructure layer agents will route millions of calls through.

Our take: a race, not a coronation

Benchmarks favor Jev, Liquid and Base Lab — but OpenAI has distribution, a zero-data-retention mode, and HIPAA eligibility, sitting inside accounts teams already have. The honest test is running your own labeled set on both. This one's a race, not a coronation.