OpenAI Releases Decisions API: Agent Decisions Up to 10x Faster

OpenAI has moved Decisions API from public beta to general availability. It is a dedicated endpoint on GPT-6 Luna that answers questions with a predefined set of answers. The company claims it does so up to 10 times faster than the same Luna through Responses API. Pricing is $0.10 per million input tokens and $0.01 per million cached tokens, and there is no charge for output tokens at all.

The idea is simple. Applications and agents make dozens of small decisions in a single workflow: which category an incoming document belongs to, which queue to send a ticket to, whether to continue, retry, or ask a follow-up question. Each decision is trivial. But when a full generative model makes it, latency and cost pile up along the chain, and an interactive product starts to lag.

Three Result Types Instead of Free-Form Text

The developer describes the question and the allowed answers, then supplies context as text, images, or native JSON. The API returns one of three result types:

• predicate: a probability from 0 to 1 that a condition is true, for example that damage is visible on a product;
• choice: one of the supplied options, plus probabilities and confidence. It suits selections such as "continue / retry / clarify";
• score: a rating on an ordered rubric scale, for example the severity of damage. The value is computed as the probability-weighted average of the level indices.

The economics follow from this design. Because the result is built as a probability distribution over the given options, the model does not need to generate text, so there is nothing to pay for on output. All the cost sits in the input. Say a classifier receives 500 tokens per call. A million calls would cost $50. If 400 of those tokens are a cacheable prefix with the instructions and rubric, the bill comes to about $14. This is a back-of-the-envelope calculation based on the published rates, but it shows the order of magnitude.

What Has Been Added Since the Beta

• native integration with GPT-Live: a decision is launched by a trigger in `session.triggers` on an event such as `session.input_transcript.delta`, and the application receives `decision.event`;
• prompt caching for repeated context, with cacheable entries lasting 30 minutes and 12 hours;
• regional inference processing in the US and EU, and Zero Data Retention for eligible customers;
• native JSON inputs and images directly in the question instructions;
• up to 400 questions per request, twice the previous limit;
• additional capacity in several Asian regions to reduce latency.

Customer Numbers

OpenAI cites five external cases. All of them come from the companies' own evaluations; there are no independent benchmarks.

• Box: document routing, with route selection ~5x faster on 30 insurance certificates at the same accuracy as Responses API. Uncertain cases go to human review.
• Datadog: assessing the risk of shell commands generated by an agent before they run. At p95 it is 40 times faster than frontier models at comparable accuracy, which let the team put Decisions in the hot path of automatic remediation.
• Ramp: attributing AI spend to actual work. It is 4 times faster and 30% cheaper than GPT-6 Luna, with no loss of accuracy.
• Sierra: choosing the next action in live customer conversations. p95 is under 280 ms, roughly 3 times faster and 90% cheaper per call than the previous approach, with no drop in outcomes for customers.
• SpeakBox: integration with GPT-Live cut the time to the next action from 1.4 s to 0.13 s, a reduction of more than 90%.

Inside OpenAI, the dots, Personal Finance (categorizing transaction descriptions), and Communications teams (tagging posts by topic and assessing whether they reinforce key narratives or challenge them) already use the API.

Where to Start

The Datadog case shows where the API is most useful: as a protective layer in front of an agent's irreversible actions. Assessing a command's risk in a fraction of a second makes it possible to check every step rather than a sample. Box's setup of handing uncertain cases to humans relies on the confidence field in the choice response.

A practical sequence looks like this. Find the one decision your application makes most often. Formulate the question and a closed set of answers. Collect examples from real traffic and run them through Decisions Playground on your own texts or images. Compare the results with your current Responses API setup on accuracy, p95, and cost. Set the probability and confidence thresholds based on labeled data, not by eye: they determine how many cases go to manual review. For voice agents, the next step is to connect the same question as a trigger in GPT-Live.