OpenAI's new Decisions API delivers fast, constrained answers

What's new

OpenAI has opened the Decisions API, a new way for companies and developers to use AI. Instead of getting a long text back, the developer asks questions with fixed answer options, and the model responds within those bounds. It launched in public beta on October 6, 2026, after a limited preview at DevDay 2026, and OpenAI's documentation now describes it as generally available.

For now, it works only with the gpt-6-luna model. There are three question types: whether a statement is true (with a probability), a choice between given options, and a score on a scale with 2-10 levels. You can ask 1-400 questions in a single call.

What's smart about it

The answers come with probabilities. That means a company can decide for itself how confident the AI must be before, for example, a case is passed on to an employee. Typical tasks include sorting inquiries, prioritizing, checking conditions and choosing an AI agent's next action.

Input can be text, images or both. According to OpenAI's documentation, images can be supplied as embedded data or as publicly accessible web addresses, and function calls and their outputs can be included as evidence. Input can also be JSON, a common data format. Files, audio and file IDs are not supported.

OpenAI says the answers come up to about 10 times faster than with the same model via the regular Responses API. That is the company's own claim. A slide from the preview, reproduced by Firecrawl, illustrated it with 150 milliseconds versus 1.6 seconds. These are not independent measurements.

Cheaper or better?

Uncached input costs $0.10 per million tokens (chunks of text), and output is free. Cached input is priced as standard gpt-6-luna caching, and it requires at least 1,024 tokens of shared input. Surcharges may apply for regional processing and very long inputs. According to the reseller LLM Gateway, it costs $0.20 per million tokens above 272,000 tokens, and the model has a context window of about 1.05 million tokens. That figure does not come from OpenAI.

Competitors are cheaper on paper. Jev from TypeSafe and Microsoft-Decision-1 both cost $0.042 per million input tokens, according to LLM Gateway. Both are text-only, and Microsoft-Decision-1 has a 32,000-token context window. According to Firecrawl, Jev has a P50 latency (typical response time) of 0.21 seconds. Firecrawl calls OpenAI's offering a constrained language model versus Jev's purpose-built model, but points out that OpenAI has image support.

There is no monthly price, because it is a tool for developers and not an app. Danish pricing has not been announced.

What it's good for

The Decisions API suits companies that need to make many small, quick decisions: sorting customer emails, prioritizing cases, assessing whether an application meets the requirements or determining what an AI agent should do next. Because images can be included, it can also be used to, for example, assess a submitted photo.

For Danish and European companies, it is relevant that regional processing is available in the US and Europe (the EEA plus Switzerland), and that data residency is available in all supported API regions. Zero Data Retention and HIPAA are supported for eligible customers. No separate Danish availability announcement has been found.