Every agentic workflow is mostly small decisions: route this ticket, classify this document, decide whether to retry. OpenAI has now carved that work out of the general-purpose API and turned it into a separate, cheap product. The Decisions API is generally available, built on GPT-6 Luna, and designed for questions with bounded answers rather than open-ended generation.
The headline number is the speed. OpenAI claims near real-time decisions up to 10x faster than GPT-6 Luna through the Responses API, which is the gap that matters when a voice agent has to choose an action between two spoken sentences. The price reflects the same specialisation: $0.10 per million input tokens, $0.01 for cached input, and — critically — no output-token charges at all. You pay for what you read, not for what the model says. For a router that reads a 2,000-token context and returns a single label, that inverts the usual economics of LLM calls.
Three Answer Shapes
The API exposes three answer types, and the distinction matters more than it sounds. A predicate returns a probability from 0 to 1 — “is this product visibly damaged?”. A choice returns one of your supplied options plus probabilities and a confidence value — “should the agent continue, retry, or ask for clarification?”. A score returns the probability-weighted average of ordered rubric levels, which is how you express severity on a scale without collapsing it to a binary.
Developers can pass native JSON inputs, embed images directly in the question instructions, and ask up to 400 questions per request — twice the previous limit. Prompt caching for reused context and native integration with GPT-Live are both in, and OpenAI added inference residency in the US and EU plus Zero Data Retention for eligible customers.
Why It Exists as a Separate Endpoint
OpenAI’s own use cases are telling. The Personal Finance team uses Decisions to categorise transaction descriptions. The Communications team scores whether coverage reinforces or challenges key narratives — a bounded judgement call that determines what gets reviewed further. The dots team uses it to cut response times. Customers are using it to attribute AI costs and to drive language-learning experiences.
The pattern is the same everywhere: a high-frequency, low-ambiguity classification sitting in the hot path of a product, where the frontier model was being asked to do a reasoning job it was overpriced for. Routing that work to a specialised endpoint is the same reasoning that produced cheap “flash” tiers — spend frontier intelligence only where the answer genuinely needs it.
The caveat is the same one that applies to any bounded-answer abstraction: a probability is only as good as the question and the context supplied with it. The API makes it cheap to ask a bad question thousands of times per minute. Teams using Decisions will need the same evaluation discipline they would apply to a model they fine-tuned themselves — and the Playground in the OpenAI platform, which accepts your own text or images, is a reasonable place to start.
Related: OpenAI Decisions API