The Per-Call Decision API Is Now a Product Category

At DevDay on Tuesday, OpenAI launched the Decisions API: a per-call endpoint that takes a question, a predefined set of answers, and some context, and returns one of those answers. No generated text. No open-ended output. Just a decision.

The API runs on GPT-6 Luna, OpenAI's low-cost model, and returns answers ten times faster than a standard Luna call. Developers define the answer options in advance, pass context as text or images, and get back the index. OpenAI lists the intended uses as content classification, request routing, and deciding an agent's next step. It launched as a limited preview at DevDay, with broad availability expected within days.

That third use case is the one to sit with. An orchestrator's routing step, the part that decides whether a task goes to a code agent, a search agent, or a retrieval agent, currently runs as a full model call. The Decisions API turns that into a structured endpoint call with a fixed answer space. It is faster, and it removes the risk of the routing model generating unexpected output that breaks the downstream handler.

TypeSafe AI's Jev, released by former OpenAI researcher Diogo Almeida's startup in mid-September, was positioned for the same problem. The Decoder reported that Jev's launch set off a wave of developer interest. OpenAI's entry roughly two weeks later is a response to demonstrated demand.

Two competing decision-oracle products in two weeks says something concrete: the routing and classification step in multi-agent systems is being treated as a separable, purchasable service rather than a design choice embedded inside the orchestrator. That is new. Six months ago, every orchestrator rolled its own routing logic. Now there is an off-the-shelf API for it and competition for who provides it.

The builder question this raises is direct: which of your orchestrator's classification steps are better served by a purpose-built decision endpoint than by a full inference call? The answer set must be fixed, which rules out open-ended tasks. But for routing, intent classification, and binary policy decisions, fixed answer sets are the right design for that step.

The Decisions API is in limited preview. Jev is generally available. For builders who want to test the pattern before OpenAI reaches full availability, both are now options.