OpenAI Launches Decisions API for Faster, Narrow AI Workflows

OpenAI says its Decisions API is now generally available through a dedicated endpoint built with GPT-6 Luna. It targets the repeated, bounded choices inside software: classify a message, select a route or decide which action an agent should take next.[1]

The company says these calls can run up to 10 times faster than GPT-6 Luna through its Responses API. That is OpenAI’s comparison, not an independent benchmark. Developers should treat the headline as a reason to test the endpoint on their own data and latency targets.[1][2]

The endpoint asks a model to choose from a constrained set of outcomes instead of composing a broad response. OpenAI documents three result types: a probability that a condition is true, a choice from values supplied by the developer, or a score against an ordered rubric. A retail system could define choices such as continue, retry or send to a person, then handle each result through its own code.[1][2]

Developers provide the question and the allowed answers, along with the information needed to make the decision. OpenAI says requests can include text, images or native JSON, and one request can contain as many as 400 questions. The response may also include confidence and probabilities for a choice. These limits describe the API contract; they do not guarantee that a model will choose correctly on every record.[1][2]

OpenAI describes the Decisions API as up to 10 times faster than GPT-6 Luna used through the Responses API. It says the focused endpoint is designed for near real-time decisions and has added prompt caching, GPT-Live integration and more capacity in several Asian regions since public beta. The announcement gives no independent test data or Malaysian latency figures.[1]

The published input rate is US$0.10 per million tokens. Cached input is listed at US$0.01 per million, and OpenAI says there is no output-token charge for this endpoint. A low per-token rate does not determine the cost of a whole workflow: developers must still count repeated context, retries, image inputs, storage, monitoring and the surrounding application.[1][2]

OpenAI’s customer examples include document classification and routing, risk scoring for agent-generated commands, and transaction categorisation. The reported speed and savings figures come from those companies’ own evaluations. They show what buyers are testing, but do not establish that the same outcomes will hold for a Malaysian SME or its language mix.[1]

The API returns a structured result for the application to interpret. A developer can keep a person in the loop when confidence is low, reject an answer that does not match an allowed value, and record which rule or model response led to the next step. OpenAI’s page gives examples, but it does not prescribe a safe design for every workflow. Those controls remain the buyer’s responsibility.[1][2]

That boundary matters when a choice can affect a payment, customer account or staff record. A team should begin with a low-risk task and compare model choices against cases a person has already reviewed. Measure the wrong-route rate as well as response time, and set a manual path for uncertain or unsupported inputs before allowing the result to trigger an external action.[2]