OpenAI Decisions API now available on AI Gateway

OpenAI's Decisions API is accessible via an OpenAI-compatible /v1/decisions endpoint on AI Gateway, allowing the use of decision models like GPT-6 Luna for routing and triage tasks instead of text generation. Decision models provide structured outputs—probabilities, choices, and scores—based on typed questions about a shared input. These models can process predicate (yes/no), choice, and score questions in a single request. Specific decision model IDs, such as openai/gpt-6-luna-decisions, are separate from standard text generation models. Access to the API is provided through the OpenAI SDK, the AI SDK, the HTTP API at /v1/evaluate, or the AI CLI. When using the SDKs, questions and answers are keyed by name, and yes/no questions use boolean types.

The introduction of a structured Decisions API shifts the utility of large language models from open-ended text generation toward deterministic, measurable classification and routing. This mechanism formalizes the use case for LLMs in operational systems like triage and guardrails by forcing probabilistic reasoning into discrete, actionable formats (choices and scores). The separation between standard generation models and decision models reflects a strategic division: one space handles creative synthesis, the other handles structured inference, which implies a growing need for systems that can reliably output quantifiable metadata rather than narrative. The requirement to use these structured outputs inherently resists the tendency of LLMs to default to verbose explanations, pushing them toward functional utility in complex workflows. This architectural distinction raises questions about the inherent tension between maximizing linguistic fluency and ensuring operational accountability when deploying AI decision-making at scale.