AI · 23 October 2026 · 5 min read
OpenAI’s Decisions API is generally available. It answers a bounded question, and the speed figure in print is a customer’s
The API, built on GPT-6 Luna, went to general availability on 22 October. OpenAI says it is up to ten times faster than calling Luna the ordinary way, and charges only for input: ten cents per million tokens, one cent if the input is cached. Sierra says its own calls come back in under 280 milliseconds. That is Sierra’s measurement, not a guarantee.
OpenAI said on 22 October that the Decisions API is generally available. It is a way to ask GPT-6 Luna a question that already has a bounded set of answers: classify this, route this, pick the next action. The point, in the company’s words, is that those small choices are simple one at a time and expensive when an application makes them all day. The API is supposed to do them from text, images and structured input, up to ten times faster than asking Luna through the ordinary Responses API. “Up to” is the ceiling, not the typical call.
The price is published, which is more than most of these notes manage. A million input tokens costs $0.10. A cached input costs $0.01. There is no charge for output tokens, because the product is not a long reply. Since the public beta, OpenAI says it has added a link to GPT-Live so a voice application can react while someone is speaking, caching for repeated context, the choice of running the inference in the United States or the European Union, native JSON, images inside a question, and up to 400 questions in one request, twice the old limit. The names it gives as users include Box, Datadog, Ramp, Sierra and Speak.
One customer’s clock
The only millisecond figure on the page is not OpenAI’s. Sierra says it uses the API to choose the next action in a live customer conversation, that those calls run at under 280 milliseconds at the 95th percentile, and that this is about three times faster and 90 percent cheaper than what it did before, with no drop in outcomes. That is Sierra describing Sierra. It is not a latency OpenAI promises every caller, and it is not a test a reader can rerun. A target of a tenth of a second, or a claim about a rival’s valuation, is not in the announcement.
What changed is that a bounded decision now has its own endpoint, its own price, and a general release rather than a beta. The model behind it is still Luna. A hard problem is still meant to go to a slower model. The split is real only if the application actually asks a question with a finite list of answers. An open-ended one will not get faster by being sent here.