OpenAI opens its Decisions API to all customers, with native voice integration

OpenAI made its Decisions API generally available on October 22, giving developers a dedicated software interface for quick judgments inside applications. The release adds native GPT-Live voice integration and more ways to include images. Its job is narrow: classify an input, route a request or choose the next action—not compose a conversational reply.

A competitor supplied the spark

The project began with OpenAI’s team looking at TypeSafe’s Jev, a model designed to make choices from supplied options. Nikunj Handa, OpenAI’s API product lead, told TechCrunch that colleagues immediately saw many uses for the approach. His team moved to offer its own version.

Rather than train a dedicated decision model, OpenAI reconfigured Luna to make inference—the processing that produces an answer—faster and cheaper. Handa said it remains the existing model underneath. OpenAI first announced the Decisions API at its September 29 developer day; the October release takes it into general availability.

Handa described the main gain as shrinking waits of one to three seconds to roughly 100–180 milliseconds. OpenAI’s own announcement claims decisions can run up to 10 times faster than GPT-6 Luna through its Responses API. That comparison is against another way of accessing Luna, not against Jev.

Preset answers, with a larger production toolkit

Developers define the question and allowed answers before submitting context as text, images or native JSON, a structured data format. The API returns a probability, a predefined choice or a score, rather than an open-ended written answer. Applications can then use that result to determine their next step.

• Probability: estimate whether a condition is true, such as whether a product has visible damage.
• Choice: select a supplied action, such as continuing, retrying or asking for clarification, with probabilities and confidence included.
• Score: rate an input against ordered levels, such as degrees of damage severity.

Since public beta, OpenAI has added prompt caching for reused context and doubled the request limit to 400 questions. Developers can also include images within question instructions. The native GPT-Live integration lets voice applications make these choices during conversations—for example, deciding whether to clarify a request or escalate an issue.

Pricing is $0.10 per million input tokens, or $0.01 for cached inputs, with no output-token charges. OpenAI also offers regional processing in the US and EU, plus Zero Data Retention for eligible customers. It says additional capacity in several Asian regions has helped improve latency.

The next test is the decision, not just the delay

Early customer examples show different ways to put that speed to work. Box says route selection was about five times faster across 30 insurance-certificate cases, with the same accuracy as the Responses API. Its workflow automates routine cases and sends lower-confidence ones to reviewers. Ramp reports four-times-faster AI cost attribution at 30% lower cost than GPT-6 Luna, without losing accuracy.

Those are customer-reported results from particular workflows. As TechCrunch notes, the useful balance of intelligence, cost, speed and calibration—how well confidence matches correctness—varies by workload. OpenAI’s launch guidance starts with one recurring decision and asks developers to test examples that reflect real usage.

Handa’s next ambition is faster still: he floated 60-millisecond responses as a way to imagine calling the API for every video frame. He also raised possible future specialized training. Neither is an announced capability of this release; today’s service still rests on reconfiguring Luna for a tightly defined job.