OpenAI’s Decisions API, announced at DevDay 2026, is a limited-preview endpoint built on a version of GPT-6 Luna that takes context plus a developer-defined set of allowed answers and returns a selection rather than generated text. Developers pass context as text or images, define a bounded question, and get back a value their code can branch on for content classification, request routing, or an agent’s next step. OpenAI says it makes decisions ten times faster than calling GPT-6 Luna through the regular API, and coverage of the launch reports broader availability expected within days.

A community guide published today on Hugging Face works through where the pattern fits and where it breaks:

  • Bounded questions only. “Which queue owns this ticket?” with four allowed answers is a decision; “write the best response” is not. Semantic contract first, unlike JSON mode, which formats text after generation.
  • A decision is a signal, not authorization. Model judgment intersected with deterministic policy produces the executable path; permissions and business policy stay in code, and a confidence score is not a calibrated probability.
  • Tool-call gates and model routing are the other uses it highlights: asking whether an action needs human confirmation, or classifying difficulty and risk before choosing a cheap model, an expensive one, retrieval, or human review.
  • Preview caveats. Endpoint naming, authentication scope, supported inputs, limits, score semantics, region availability, and pricing may still change, so verify the official schema and keep the integration behind a small adapter rather than shipping against a third-party paraphrase.

Related: TypeSafe AI Introduces System One Models and Jev, OpenRouter Picks TypeSafe’s Jev to Route Model Selection, OpenAI Releases GPT-6 Sol and GPT-6 Luna

What Is OpenAI Decisions API? A Practical Guide · The Decoder: OpenAI expands Codex and its API at DevDay