OpenAI made its Decisions API generally available on October 22. The service answers short, bounded questions for apps and AI agents, and it runs on GPT-6 Luna.

Pricing is the headline. OpenAI charges $0.10 per million input tokens, or $0.01 per million for cached input. There is no charge for output tokens.

The company says Decisions returns answers up to 10 times faster than sending the same model through its general Responses API. It works with text, images and native JSON.

Since the public beta, OpenAI has added prompt caching, US and EU inference residency, a native link to its GPT-Live voice product, and a higher cap of 400 questions per request, double the earlier limit. Details are in OpenAI’s announcement.

Key takeaways

What is the OpenAI Decisions API used for?

It handles the small calls inside a workflow: classify a document, pick a route, score a risk, choose the next step. A developer writes the question, lists the allowed answers and passes in context.

Three answer types are supported. A predicate returns a probability between 0 and 1. A choice returns one of the developer’s own options with a confidence figure. A score grades input against a rubric of ordered levels.

OpenAI published customer figures with the launch. Box reported about 5x faster routing on 30 insurance-certificate cases at the same accuracy. Datadog cited 40x faster p95 latency when scoring agent-generated shell commands. Sierra said it runs under 280 milliseconds at p95 in live customer conversations.

Ramp reported cost attribution that was 4x faster and 30% cheaper than calling GPT-6 Luna directly. Speak said time to next action in a voice lesson fell from 1.4 seconds to 0.13 seconds.

The design separates cheap, repeated judgment calls from the expensive generation step. For creators and small teams wiring agents into publishing pipelines, that split is the practical change; more on that layer in How AI is Reshaping the Content Creator Stack in 2026.

The release landed the same week OpenAI models entered Google Cloud’s Model Garden, as reported by TechTarget. WebProNews reported that GPT-6 Luna was among the models listed there in private preview.

What to watch next

Decisions supports only GPT-6 Luna for now, according to WebProNews’ reading of OpenAI’s developer guide. Support for a second model, or published accuracy benchmarks against the Responses API, would show how far OpenAI plans to push the endpoint beyond routing and labeling.