OpenAI’s Decisions API beats its own engine by up to 10x

OpenAI opened its Decisions API to all developers in public beta on October 6, which ends the limited preview, the company said in a post on the OpenAI developer forum. Classification and routing calls that go to GPT-6 Luna through the Responses API today can move to the new endpoint on the same model, which OpenAI says returns decisions up to 10x faster. The limit is that gpt-6-luna is the only model it currently serves.
The API doesn’t hold a conversation. It takes text and image inputs and returns one of three output types:
- Predicates: the probability that a statement is true.
- Choices: a pick from options you define in advance, with confidence scores.
- Scores: a score for an input you ask it to evaluate.
OpenAI pitches it for apps that need to choose the right model, tool or action in near real time, such as a router in front of several models or an agent picking its next step.
The 10x figure is OpenAI’s own, and the baseline is Luna itself, called through the Responses API, not some other model. When the API was still in limited preview in September, OpenAI gave latency figures of 150 milliseconds per decision against 1.6 seconds for GPT-6 Luna. Those are the company’s numbers from the preview, and the beta post repeats only the “up to” multiple.
Luna launched on September 22 alongside GPT-6 Sol, and OpenAI describes it as suited to “high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions,” TechCrunch reported. That is the kind of work the Decisions API is built for, and Luna is still the only engine behind it: a decision that needs a stronger model has nowhere to escalate inside this API.
A trained classifier needs retraining when its labels change. In the preview, OpenAI’s answer was to accept new labels right in the prompt, with no labeled data and no retraining, so changing a routing table means editing the list of choices you send.
Up next
OpenAI’s math results cost about three hours of Pro each
An unreleased OpenAI model spent roughly three hours of ChatGPT Pro thinking, on average, on each new math result it produced, and over the evaluation it was posed approximately 4,000 problems.
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

