OpenAI opens Decisions API, prices input at $0.10 per million tokens

Key Points

1. OpenAI released its Decisions API publicly, built on GPT-6 Luna for constrained, fast machine decisions.
2. The interface returns probabilities, preset choices with confidence scores, or probability-weighted scores from a rubric.
3. OpenAI says it runs up to 10 times faster than calling the same model conventionally.

The latest:

OpenAI has made its Decisions API generally available, a GPT-6 Luna interface built for constrained outputs such as classification, routing and next-action selection. The company says the endpoint is up to 10 times faster than calling GPT-6 Luna through the Responses API, and prices it at $0.10 per million input tokens with no charge on output tokens.

Details:

• The three outputs: According to OpenAI, the API returns one of three response types: a Predicate, a probability between 0 and 1 that a condition holds; a Choice, a selection from predefined values with confidence scores; and a Score, a probability-weighted rating against a scoring rubric. The design restricts the model to bounded answers rather than open text.
• Pricing: OpenAI lists input at $0.10 per million tokens and cached tokens at $0.01 per million, with output tokens billed at nothing. That structure shifts cost entirely to what developers send in, favoring high-volume, repetitive decision calls where prompts are largely reused and cached.
• Inputs supported: The company says the interface accepts text, images and structured JSON data, allowing a single call to weigh a customer record, a document scan and a written request together rather than requiring separate passes through different endpoints.
• Voice integration: Decisions API integrates directly with GPT-Live for real-time voice applications, per OpenAI, handling in-call judgments such as whether to ask a clarifying question or reroute a request while the conversation is still running.
• Capacity raised: OpenAI doubled the per-request ceiling to 400 questions, up from the limit in the earlier preview release, meaning a single call can resolve hundreds of separate classifications at once.
• Customer figures: Among the performance examples OpenAI cited, Sierra recorded 95th-percentile latency under 280 milliseconds, roughly 3 times faster and 90% cheaper per call than its previous method, and Speak cut time-to-next-action in GPT-Live tests from 1.4 seconds to 0.13 seconds.
• Datadog benchmark: Datadog logged 95th-percentile latency 40 times faster than available frontier models at comparable accuracy, according to figures in OpenAI’s announcement. The company did not publish the underlying test methodology or the accuracy threshold used for that comparison.
• Data controls: OpenAI says inference data residency is supported in the United States and the European Union, with a Zero Data Retention option available to eligible customers — terms aimed at regulated buyers in finance, healthcare and public-sector work.
• The positioning: By pricing output at zero and capping the response format, OpenAI is selling the endpoint as infrastructure rather than a chat model: a component called thousands of times inside an application flow, not a product an end user talks to.

Between the lines:

The pricing and the latency figures point at the same buyer. Zero-cost output tokens and $0.01 cached input only matter at volume, and sub-280-millisecond responses only matter inside live systems. Together with the GPT-Live integration and the jump to 400 questions per request, the release targets companies replacing rule engines and smaller classifier models inside production pipelines.

What’s next

Watch whether the cited latency gains hold outside OpenAI’s selected customer examples, and whether rival providers answer the zero-cost output pricing. Independent benchmarks against frontier models at matched accuracy would test the Datadog claim.