OpenAI took its Decisions API out of beta on October 22, 2026, pricing it at $0.10 per million input tokens, with no output-token fees and cached input at $0.01 per million. The endpoint, POST /v1/decisions, runs on GPT-6 Luna and returns either a 0-to-1 predicate or a two-to-ten-level ordered score, up to 400 questions per request, doubled from the beta’s 200.

The pitch is latency. OpenAI says the endpoint is up to 10x faster than routing the same question through the Responses API, where Luna output alone costs $0.50 per million tokens. Nikunj Handa, OpenAI’s API product lead, told TechCrunch the team didn’t train a bespoke classifier for this: “it’s just plain old Luna that gets the job done.” The “big unlock,” he said, is dropping typical next-action latency from one-to-three seconds to 100-to-180 milliseconds.

The design-partner numbers are the real story. Speak’s CTO Andrew Hsu says time between actions fell from 1.4 seconds to 0.13 seconds, a reduction of more than 90%. Sierra’s Ajeet Grewal reports under 280ms p95 next-action latency, roughly 3x faster at 90% lower cost per call. Datadog’s AJ Stuyvenberg clocked 40x faster p95 latency scoring agent-generated shell commands against available frontier models. Ramp’s René Sultan put AI cost attribution at 4x faster and 30% cheaper than calling Luna directly. Box’s Biswaroop Palit measured about 5x faster route selection across 30 certificate-of-insurance cases at the same accuracy as the Responses API. Prompt caching kicks in at prefixes of at least 1,024 tokens, with cache reads at $0.01 per million.

For a five-to-thirty-person B2B team, the use case writes itself: inbound-lead triage. A score-and-route call under 200ms means a web form, a chat widget, or an inbox webhook can classify intent, urgency, and fit before the prospect’s tab goes stale. Speed-to-lead has always been the quiet determinant of conversion, and the price point makes it a line item rather than a project. That connects directly to the pattern of always-on Dots assistants and the Claude Duet shoulder-tap model: the decision layer is becoming a utility, and the follow-up is the product.

Inference residency is live in the US and EU, with Zero Data Retention and HIPAA use available to eligible customers. Classifiers used to be a project. Now they’re a price list.

Sources