OpenAI announces a Decisions API on GPT-6 Luna, in limited preview
At DevDay on 29 September 2026, OpenAI announced the OpenAI Decisions API. Per the official recap, it “enables real-time decision-making by focusing Luna’s intelligence on a specific set of user-defined questions with finite pre-defined answers”. Developers send text or images as context and get back answers they can use to classify content, route requests or choose an agent’s next action.
What OpenAI said. @OpenAIDevs (15:34 BRT) calls it “powered by GPT-6 Luna” and “available in limited preview”. A follow-up says the API “returns a selection your app can use” and that preview access “is limited to selected API customers for testing”, with a broad release planned “in the coming days”. Earlier, Thibault Sottiaux (14:27 BRT) said it “supports visual inputs” and is “tuned to be able to make decisions in less than a few hundreds of milliseconds end to end”.
Vendor numbers (UnverifiedClaim). The 29-second launch video routes 10,000 customer requests to Billing / Technical / Sales and shows 150 ms per request for the Decisions API against 1.6 s on the Responses API, captioned “~10x faster decisions” and “Shown 15x realtime”. The Decoder and The New Stack repeat those figures. The New Stack also writes that the API returns “confidence scores”; OpenAI’s own posts don’t say that.
What is still unknown. There is no public documentation: no guide, no API reference, no changelog entry and no endpoint on the GPT-6 Luna model page. Pricing, the request and response schema, whether probabilities come back and are calibrated, how many answers a question can take, context limits, regions and p50/p99 latency are all unpublished. An OpenAI spokesperson told The New Stack the company will share more “at broad rollout”. GPT-6 Luna itself is a general text model, released 22 September at $0.10 / $0.50 per 1M tokens; OpenAI has not said whether the Decisions API is billed the same way.
Where it sits in the category. OpenAI doesn’t use the “System One” name or TypeSafe’s /v1/systemone interface, but press read it as a reply to Jev: “OpenAI answers TypeSafe’s Jev” (The New Stack), and OpenAI “is moving into the field of so-called System One models” (The Decoder). TypeSafe founder Diogo Almeida answered with a “clone war” meme and then: “I love openai and think more competition and validation is great for developers! (assuming the model is good - plz make it good!)” He added that he hopes it signals “building in a system one compatible way is the future” (@CompleteSkeptic, 15:24 BRT). The same day he pointed developers to TypeSafe’s system-one-adapter-python, which puts ordinary LLM APIs (OpenAI included) behind TypeSafe’s System One client interface for side-by-side comparisons.
How we list it. No decision-specific weights, model card or API contract has been published, so the OpenAI Decisions API goes under runtimes, off the quadrant. We will re-check at broad release; if OpenAI documents tuned decision weights and a typed schema, it moves to the catalog. Not to be confused with OpenRouter’s Decisions API (/api/alpha/decisions), a different product that serves Jev and other decision models.