OpenAI puts Decisions API for predefined agent actions into limited preview
AI agents · Friday, 2 October 2026
Why it matters
A structured decision layer could help agent platforms separate high-volume routing and policy checks from expensive general-purpose generation, making it more practical to gate or evaluate agent actions. Because the API is still in limited preview and lacks public benchmarks, builders should treat its reliability and cost advantages as unverified rather than as established performance.
What happened
OpenAI introduced the Decisions API at DevDay 2026, using its Luna model to select from developer-defined answers for classification, request routing, and AI-agent action selection instead of generating open-ended text. OpenAI says the approach is intended to improve speed and cost while retaining image understanding, multilingual support, and safety protections. The API remains in limited preview, and OpenAI has not yet published detailed public benchmarks for its accuracy or probability calibration.
Previously
Sep 2026 — OpenAI made GPT-6 Astra available to a limited group of users for computer use, browsing, software engineering, cybersecurity, science, and professional work.(CIO)
Players & places
- OpenAI
- TypeSafe AI