OpenAI's Decisions API Answers Quick Yes/No Questions for $0.10 per Million Tokens
OpenAI's generally available Decisions API, built on GPT-6 Luna, returns probabilities for bounded questions such as classifying or routing, and bills no output tokens.
OpenAI has made its Decisions API generally available. It is a special endpoint, built on the GPT-6 Luna model, for questions where the answer is limited to a fixed set of options: is this photo damaged, which team should handle this email, what is the next step in this workflow. Instead of writing a paragraph, it returns a number.
You define the question and the allowed answers, then supply context as text, images or JSON. There are three answer types. A “predicate” gives a probability from 0 to 1 that something is true. A “choice” picks one of your options and reports probabilities and confidence. A “score” rates something against ordered levels, such as a quality rubric, as a probability-weighted average. OpenAI says it is up to 10 times faster than using GPT-6 Luna through the normal Responses API. Customers quoted include Datadog, reporting 40 times faster p95 latency (the speed that 95 percent of requests beat), and Speak, which cut time to the next action from 1.4 seconds to 0.13.
The pricing is notable: $0.10 per million input tokens (tokens are the small word chunks a model reads), $0.01 per million cached input tokens, and no charge for output tokens. Since the output is only a short result, that makes sense. Newer additions include US and EU data residency, Zero Data Retention for eligible customers, prompt caching, images in instructions and up to 400 questions in one request. The speed and customer figures are OpenAI’s own claims.
What’s behind it: most real software does not need an AI that chats. It needs a quick, cheap judgment call thousands of times a day: spam or not, urgent or not. Using a full chat model for that is like hiring a novelist to tick boxes. A dedicated endpoint that returns probabilities is easier to plug into normal code, because you can set your own threshold, for example “only act if above 90 percent”.
What this means for you: if you do not write code, this is mostly invisible; you may just notice apps getting snappier. If you build tools or automations, it is worth testing for sorting, routing and quality checks, with a person still reviewing cases where the model is unsure. Confidence numbers are useful, but not a guarantee.
Sources
Flint: A 40B Open Coding Model That an AI System Helped Build in Four Months
Recursive's Flint is an Apache 2.0 open-weight coding model that runs on laptops and was developed largely by the company's own AI research system, though the claims are unverified.
