OpenAI Makes Decisions API Generally Available for Agents
OpenAI has released a low-cost endpoint for bounded choices, aiming to make real-time routing and agent control faster and cheaper.
OpenAI has made its Decisions API generally available, introducing a dedicated endpoint for applications that need fast, bounded judgments rather than open-ended text generation.
A narrower model interface
The service is powered by GPT-6 Luna and accepts text, images and structured JSON. Developers define the possible decision space, then request one of three outputs: a probability, a choice from a supplied list, or a scored assessment. OpenAI says the endpoint can operate up to 10 times faster than using GPT-6 Luna through the general Responses API.
The release adds native integration with GPT-Live, allowing voice applications to decide whether to clarify, escalate or take an action while a conversation is in progress. It also supports prompt caching, US and EU inference residency, native JSON inputs and batches of up to 400 questions.
Economics matter more than the demo
OpenAI prices the service at $0.10 per million input tokens, with cached input at $0.01 per million. There is no output-token charge. Those economics target the large number of small classifications that sit inside production systems: document routing, transaction categorisation, quality checks, moderation decisions and tool selection.
The deeper move is architectural. OpenAI is separating deliberative generation from repetitive control logic, giving developers a cheaper primitive for the decisions that determine whether an agent continues, retries, asks for help or stops.
Why it matters
Agent deployments often fail economically because every minor branch invokes a full language-model response. A typed decision endpoint could reduce that overhead and make voice, customer-service and workflow agents more responsive. The unresolved question is calibration: a fast probability or choice is useful only when teams can measure its error costs and build reliable escalation paths around it.
Uncle Cat take: OpenAI is pricing the control plane, not just the conversation; $0.10 per million input tokens could make tiny agent decisions economically invisible.