The OpenAI Decisions API became generally available on 22 October 2026, and it does one narrow job: it answers questions that have a fixed set of possible answers. Is this photo of a parcel showing damage? Which team should this support ticket go to? Should a voice assistant ask a follow-up question or hand the call to a person? OpenAI says the endpoint returns those answers up to ten times faster than sending the same prompt to GPT-6 Luna through its general Responses API.
That sounds like a small product. For anyone building agents or customer-facing apps, it may not be. Most of the waiting in an AI app comes from dozens of small decisions made before the visible answer appears, and this is a tool aimed squarely at those.
What the OpenAI Decisions API actually does
A developer writes a question, lists the answers that are allowed, and passes in context. That context can be text, images or structured JSON. According to OpenAI's announcement, the API then returns one of three kinds of answer:
- Predicate: a probability between 0 and 1 that a condition is true, such as "the product in this photo is damaged".
- Choice: one of the options the developer defined, with probabilities and a confidence figure.
- Score: a probability-weighted position on a rubric, for example a damage severity scale.
The model behind it is GPT-6 Luna. OpenAI did not build a separate decision model for this. Nikunj Handa, OpenAI's API product lead, told TechCrunch that "right now, it's just plain old Luna that gets the job done", with the system tuned for speed and cost on short, bounded outputs.
How fast is it, and what does it cost?
TechCrunch's Tim Fernholz reported that response times fell from about one to three seconds per request during testing to roughly 100 to 180 milliseconds now. Handa said he hopes to get to about 60 milliseconds, which he said would make per-frame decisions on video calls possible.
Pricing is simple. OpenAI lists $0.10 per million input tokens and $0.01 per million cached input tokens, with no charge for output tokens. Cache writes kept for 30 minutes or 12 hours are priced separately in OpenAI's pricing guide. Because answers are short labels or probabilities, the lack of an output charge matters less than the low input price, but it does make costs easy to predict.
Customer numbers OpenAI published
OpenAI shared results from early users. These are the companies' own figures, reported by OpenAI:
- Datadog: 40 times faster at the 95th percentile when risk-scoring shell commands generated by AI agents, with comparable accuracy.
- Box: about five times faster at routing certificate-of-insurance documents, across 30 test cases, with the same accuracy as the Responses API.
- Ramp: four times faster and 30% cheaper than GPT-6 Luna for attributing AI costs to pieces of work.
- Sierra: under 280 milliseconds at the 95th percentile, about three times faster and 90% cheaper per call than its previous setup.
- Speak: time to the next action in a voice lesson fell from 1.4 seconds to 0.13 seconds, according to its CTO Andrew Hsu.
What changed since the beta
OpenAI first showed the product at its developer day on 29 September. The general release adds a few things developers asked for:
- Native integration with GPT-Live, OpenAI's real-time voice API, so a voice session can trigger a decision on each new piece of transcript.
- Prompt caching for context that gets reused across many questions.
- Inference residency in the US and the EU, and Zero Data Retention for eligible customers.
- Images and native JSON inside question instructions.
- Up to 400 questions in one request, double the earlier limit.
- Extra capacity in several Asian regions to cut latency.
OpenAI's announcement does not list African regions. Developers in Lagos, Nairobi or Johannesburg can call the API like any other OpenAI endpoint, but requests will be processed in the regions OpenAI names, so network distance will add to the latency figures above.
Where it fits: decision models are now a category
OpenAI is not alone here. TechCrunch notes that the startup TypeSafe, valued at about $7.6 billion according to the report, sells a decision model called Jev that is the most queried decision model on OpenRouter. On two public rankings cited by TechCrunch, Jev sits 7th on the AIM System One Index and 3rd on the Artificial Analysis Decision Index, while OpenAI's offering is 8th and 14th. Liquid's D1 Ultra, released last week, and Base Lab's Elephant-1 also score well.
TypeSafe CEO Diogo Almeida told TechCrunch he distrusts benchmarks and urged developers to test models on their own work. That is sensible advice for anyone choosing between these tools, especially since the cheapest per-token option is not always the cheapest once retries and wrong answers are counted.
We have tracked a similar push for speed elsewhere, including Inception's Mercury 3 diffusion model, which competes on raw tokens per second rather than on structured answers.
Should you use it?
The Decisions API makes most sense when an app asks the same kind of closed question many times: routing, tagging, moderation checks, approving or blocking an agent's next step. In those cases a fixed answer set and a probability are more useful than free text, and the speed gain compounds across every step.
It is less useful for anything open-ended. If the answer needs to be explained, summarised or written out, a normal chat or responses call is still the right tool.
A practical way to start is OpenAI's Decisions Playground at platform.openai.com/decisions, where you can paste your own text or images and see the probabilities before writing code. Run your real cases through it, compare the results with your current prompts, and only then decide whether the speed is worth moving a workflow over.
