OpenAI makes its Decisions API generally available

By VK

Published 25 Oct 2026, 11:44 UTC

Reporting from TechCrunch, OpenAI

OpenAI's rival to Jev is open to all API customers at $0.10 per million input tokens, with no output charge and up to 400 questions per call.

What it means for founders

Price the swap before you make it. At $0.10 per million input tokens and no output charge, routing, triage and tagging calls now sent to a general model could cost a small slice of today's bill. Check accuracy on your own data first: every speedup OpenAI published comes from customers it chose to quote.
Rankings will not settle the choice. Public indexes put OpenAI behind Jev and at least two other models, while OpenAI leans on voice and regional residency, and the newest Jev handles images too. Run one decision set through two providers and keep switching cheap.
Voice products gain a lever. If you build on GPT-Live, decisions can now fire inside the session instead of through a separate classifier, which should trim the pauses users notice.
Watch the latency target. Handa floated getting calls down to about 60 milliseconds, which he suggested could allow a call on every video frame. That, or a dedicated Luna variant, would widen what these calls can do.

The story

OpenAI made its Decisions API generally available on Thursday, October 22, opening to all API customers a product first shown at its September 29 developer day. It uses GPT-6 Luna for the small, closed-ended calls an app or agent makes constantly, like sorting a ticket or choosing an agent's next move. OpenAI's announcement says it answers up to 10 times faster than sending the same work to Luna through the Responses API. It is OpenAI's answer to TypeSafe's Jev.

What changed since the beta

The general release wires Decisions directly into GPT-Live, OpenAI's voice product, so a spoken conversation can escalate a support case or make an on-screen avatar smile while the user talks. Other changes since the September beta: the cap per request doubles to 400 questions, reused context can be cached, questions can carry JSON and embedded images, inference can stay in the US or EU, and added Asian capacity trims latency.

Answers come back in three shapes: a probability that something is true, one option from a developer's list with confidence attached, or a rubric score.

Only input is billed, at $0.10 per million tokens or $0.01 when cached, and eligible customers can get Zero Data Retention.

What early users report

OpenAI's post quotes customers it selected. Speak says time to next action in its GPT-Live testing fell from 1.4 seconds to 0.13 seconds. Ramp reports cost attribution running four times faster and 30% cheaper than plain GPT-6 Luna. Sierra cites under 280 milliseconds at p95, with each call costing roughly a tenth of its old method, and Datadog risk-checks its agents' shell commands with it before they execute.

According to TechCrunch, OpenAI skipped training a dedicated model and reworked Luna's inference for speed and cost; Nikunj Handa, who leads product for OpenAI's APIs, said that for now it is "just plain old Luna."

On the public benchmarks TechCrunch cites, OpenAI's model ranks eighth on the AIM System One Index and fourteenth on the Artificial Analysis Decision Index, behind the latest Jev at seventh and third, with Liquid's D1 Ultra and Base Lab's Elephant-1 generally on top. Jev has spread fast, reaching 35% of Vercel paid teams in two weeks.

What we don't know yet

Whether OpenAI will train a purpose-built decision model, which Handa described only as a future possibility.
How the quoted speedups hold up beyond the companies OpenAI picked, since none came with a public test set.
When image outputs, which Handa said customers are asking for, might arrive.