OpenAI announced on October 22, 2026 that the Decisions API is generally available. Built on GPT-6 Luna, the API is meant for bounded answers: classify text or images, evaluate an input, and decide what happens next without generating a full model response. The company blog is titled "Build more responsive apps with the Decisions API."
Key Facts
| Details | |
|---|---|
| Product | Decisions API (generally available) |
| Announced | October 22, 2026 (OpenAI blog) |
| Model | GPT-6 Luna |
| Answer types | predicate (probability), choice, score |
| Speed claim (OpenAI) | Up to about 10x faster than GPT-6 Luna via the Responses API |
| Pricing (official) | Input $0.10 / million tokens; cached input $0.01 / million; output free (input only) |
| Per-request questions | Up to 400 (twice the public-beta limit) |
| GA additions vs public beta | Native GPT-Live integration, prompt caching, US/EU inference residency, native JSON input, images in question instructions |
| Earlier timeline | Announced at Dev Day on September 29, 2026; public beta in early October |
Background
Many agent and app workflows need many small decisions: route a request, classify a document, or pick the next action. Calling a full generative endpoint for each step adds latency and cost. OpenAI positioned Decisions as a dedicated endpoint for those bounded questions.
TechCrunch (Tim Fernholz, October 22, 2026) framed the launch as OpenAI entering a decision-model category that TypeSafe's Jev model helped popularize. Nikunj Handa, OpenAI's API product lead, told TechCrunch that latency moved from about 1–3 seconds to roughly 100–180 milliseconds, and that the public release added voice and vision modalities. OpenAI said Decisions was announced at Dev Day on September 29 and opened to all customers on Thursday after an early-October public beta.
Reactions and Impact
OpenAI published customer quotes with measured deltas. Speak CTO Andrew Hsu said time to next action fell from 1.4 seconds to 0.13 seconds in GPT-Live testing. Box's Biswaroop Palit cited about 5x faster route selection on 30 certificate-of-insurance cases versus the Responses API at similar accuracy. Datadog's AJ Stuyvenberg reported about 40x faster p95 versus available frontier models for scoring agent-generated remediation commands. Ramp's René Sultan said cost attribution was 4x faster and 30% cheaper versus GPT-6 Luna. Sierra's Ajeet Grewal said orchestration decisions ran under 280 ms p95, about 3x faster and 90% lower cost per call than their previous approach.
Independent third-party benchmarks beyond those vendor-cited figures were still thin at writing time.
What We Don't Know Yet
- How Decisions compares with TypeSafe Jev and other decision models on the same public suites (OpenAI and TechCrunch discuss the category; they do not publish a shared leaderboard in these pieces).
- Long-term pricing for 30-minute and 12-hour cache writes beyond the published per-token input rates.
- Whether more models beyond GPT-6 Luna will support the Decisions endpoint.