OpenAI makes its Decisions API generally available on GPT-6 Luna
OpenAI took the Decisions API out of beta on October 22, making it generally available on GPT-6 Luna through a dedicated endpoint. The API targets a narrow job: questions with bounded answers, the small classification, routing and scoring calls that an application or agent makes dozens of times inside a workflow. OpenAI says it answers those questions up to 10x faster than GPT-6 Luna through the Responses API, working from text, images and structured input.
The bet behind the product is simple. Individually these calls are trivial. At scale, their latency and cost add up faster than any single large model call.
Three answer types
Developers define a question and its allowed answers, then pass context as text, images or native JSON. The API returns one of three answer types. A predicate returns a probability between 0 and 1 that a condition holds, such as whether a product photo shows visible damage. A choice returns one of the values you supply, along with per-choice probabilities and a confidence score. A score rates an input against an ordered rubric and returns the probability-weighted average of the level indices.
Since the public beta, OpenAI has added prompt caching for reused context, native JSON inputs, images inside question instructions, and a per-request question limit doubled to 400. Inference residency is available in the US and the EU, and the company says it added capacity in several regions of Asia to bring latency down further.
The customer numbers, attributed
Most of the performance figures in the announcement come from customers measuring their own workloads, and they are specific. Box says the API cut route-selection time about 5x across 30 certificate-of-insurance cases in its evaluation, with the same accuracy as the Responses API. Datadog reports 40x faster p95 latency than the frontier models it compares against, and uses Decisions to score the risk of agent-generated shell commands before they run. Ramp says cost attribution for its AI usage runs 4x faster and 30% cheaper than GPT-6 Luna, with accuracy unchanged. Sierra puts p95 under 280 ms, roughly 3x faster and 90% lower cost per call than its previous approach. Speak, which builds language-learning apps, reports that the GPT-Live integration cut time to the next action by over 90%, from 1.4 seconds to 0.13 seconds.
None of these numbers has been independently reproduced yet. Each is a company's measurement on its own traffic, which is the honest way to read the list.
OpenAI runs it internally as well. The dots team uses it to cut response times in its app, the Personal Finance team categorizes transaction descriptions, and the Communications team classifies news coverage by theme and scores whether it reinforces or challenges key narratives.
Voice: Decisions inside GPT-Live
The native GPT-Live integration is the newest addition. During a live voice session, transcript updates can trigger a decision, and the session emits a decision event that the application reads. OpenAI's example picks an avatar's facial expression from the user's latest message. A support agent could use the same mechanism to decide whether to ask a clarifying question, route a request or escalate an issue.
Pricing, and who should care
Decisions API costs $0.10 per million input tokens, and cached input costs $0.01 per million. There are no output-token charges at all, which matters when the answer is a single number or a single choice. Cached writes come in 30-minute and 12-hour tiers, and Zero Data Retention is available to eligible customers.
The competitive context is the story. Vercel's Jev, the decision-focused model this site covered in October, reached 35% of the gateway's paid teams in two weeks. Liquid AI's d1-ultra markets itself as a decision model that returns probabilities instead of tokens. OpenAI entering that lane with a dedicated endpoint and input-only pricing makes it a real option in the category now.
If your app makes high-volume small calls, classification and routing above all, the practical move is to point one recurring decision at the Decisions Playground, measure it against your current model on your own examples, and only then route production traffic. The speed and cost figures in the announcement are OpenAI's and its customers' own. Your workload is the one that has to clear the bar.
Source: openai.com
More news
Have something in mind?
Start a conversation