OpenAI has made its Decisions API generally available, offering developers a faster and more cost-effective way to handle bounded decisions in applications and agents. Built with GPT-6 Luna, the API focuses the model's intelligence on questions with defined answers, such as classifying inputs, routing requests, or selecting next actions. The company says the service can operate up to 10 times faster than GPT-6 Luna through the Responses API while supporting text, images, and structured inputs.
What the Decisions API Offers
The Decisions API lets developers define a question and its allowed answers, then provide context as text, images, or native JSON. It returns one of three answer types: predicate, choice, and score, each designed for specific workflows. A predicate estimates the probability that a condition is true, a choice selects from predefined options, and a score rates an input against ordered rubric levels.
Capabilities Added Since Beta
Since the public beta, OpenAI has added native GPT-Live integration for real-time voice experiences, prompt caching for reused context, and inference residency in the United States and European Union. Developers can also pass native JSON inputs, include images in question instructions, and ask up to 400 questions per request, double the previous limit. The company has continued to improve latency and added capacity in several Asian regions.
Internal Applications at OpenAI
OpenAI reports that teams across its own organization are using the Decisions API in product and business workflows. The Dots voice assistant team uses it to reduce response times, while the Personal Finance team categorizes transaction descriptions. The Communications team applies decisions to classify coverage by themes and assess whether it reinforces or challenges key narratives.
Enterprise Customer Adoption
External customers are also adopting the API for diverse use cases. Box uses Box Extract and the Decisions API to classify business documents and route them to the right next step, reporting approximately 5 times faster route selection with the same accuracy as the Responses API. Speak, which builds AI language tutors, reduced the time between actions from 1.4 seconds to 0.13 seconds, while Datadog, Ramp, and Sierra also participated in early trials.
The Rise of Decision Models
The launch places OpenAI in a growing market for fast and inexpensive decision models popularized by TypeSafe's Jev model. TypeSafe, founded by a former OpenAI researcher, became a US$7.6 billion company by creating deep learning models that act as super classifiers, giving choices and confidence ratings. OpenAI product lead Nikunj Handa said his team reconfigured ChatGPT Luna to accelerate inference and lower cost rather than building a separately trained decision model.
Speed and Benchmark Positioning
Handa noted that the main unlock has been speed, shrinking typical response times from 1 to 3 seconds down to 100 to 180 milliseconds. He also suggested that future versions could receive specific training for decision tasks, but current performance relies on Luna. Public benchmarks such as the AIM System One Index and the Artificial Analysis Decision Index currently favor D1 Ultra and Base Lab's Elephant-1, with OpenAI's model placing eighth and fourteenth respectively.
Multimodality and the Road Ahead
OpenAI sees multimodal capabilities as a key advantage for its decision offering. Handa told TechCrunch that Luna is extremely powerful for image classification and has shown better calibrations, and customers want more flexibility with images as outputs. He also sees further speed improvements enabling new use cases, potentially reaching 60 milliseconds so that every frame of a video could trigger an API call.
As decision models become faster and cheaper, OpenAI expects adoption to expand into a wider set of use cases that previously remained impractical. The company plans to continue improving speed, calibration, and flexibility while monitoring how developers apply the Decisions API across voice, vision, and structured workflows. For now, the general availability of the API gives builders a new tool to turn incoming information into useful action with lower latency and cost.