OpenAI’s Decisions API leaves beta, with caching and European processing

More context, more questions

The current developer guide confirms support for text, images and native JSON input, including images inside individual question instructions. A request can now ask up to 400 questions about shared evidence, twice the beta limit. Caching allows repeated context to be reused across requests.

OpenAI also added a native GPT-Live integration, so a voice application can trigger decisions as conversation transcripts arrive. The company’s announcement below introduces the generally available release; its embedded demonstration comes from the earlier beta.

A small answer still needs a test

The output is deliberately constrained. A choice selects from supplied categories; a predicate estimates whether a condition is true; a score rates an input against ordered levels. A support queue could use those answers to sort messages and flag cases for review. That is an example of the documented design, rather than a reported deployment.

OpenAI claims decisions can run up to ten times faster than calling Luna through the Responses API. Its launch includes customer-reported improvements, but those selected accounts do not establish a general speed or accuracy advantage. A team still needs labelled examples and thresholds that reflect the cost of routing a message incorrectly.

Our independent-test coverage of Microsoft Decision-1 shows why probabilities deserve their own check. That evaluation concerns Microsoft’s model; it is not evidence of how this OpenAI endpoint performs.

The bill and the region are separate settings

Uncached input costs $0.10 per million tokens, with no output-token charge. Cached reads cost $0.01 per million; cache writes are charged at the standard Luna rates shown below. Regional-processing premiums and long-context multipliers still apply, so “input only” does not mean every input token has the same price.

For European teams, the data-controls guide distinguishes availability from processing location. Decisions is available across supported API regions, while regional inference is supported in the US and Europe, defined as the EEA plus Switzerland. Non-US data residency requires approval for abuse-monitoring controls and a Modified Retention amendment. Zero Data Retention is available to eligible customers. Selecting a European project is part of a deployment configuration, not a substitute for those requirements.