OpenAI made its Decisions API generally available on October 22, 2026, offering developers a specialized tool for fast, narrowly defined choices inside applications and automated workflows. Built with GPT-6 Luna, the service targets tasks such as classifying documents, routing requests and choosing an agent’s next action, with charges based only on input tokens.
Key Takeaways
- OpenAI says the service can make bounded decisions up to 10 times faster than GPT-6 Luna through the Responses API.
- Box and Datadog reported faster performance in specific evaluations, rather than a universal benchmark for every application.
- The launch adds native GPT-Live integration, prompt caching and regional processing in the US and EU.
Table of Contents
- Decisions API targets small choices inside larger workflows
- Customer tests show gains, with different comparisons
- Pricing and deployment options shape implementation
- Conclusion
Decisions API targets small choices inside larger workflows
OpenAI’s general availability announcement positions the product around recurring decisions whose possible answers are defined in advance. Developers supply a question, allowed answers and context, which can include text, images or native JSON.
The service returns a probability, a choice or a score rather than an open-ended response. That makes it suited to selecting a workflow branch, assessing whether a condition is true or rating an input against a defined rubric.
The practical distinction is between generating a broad answer and making a constrained choice that another application can act on. OpenAI says the latency and cost of these individually small decisions can accumulate when applications repeat them at scale.
| Item | Details |
|---|---|
| Standard input price | $0.10 per million input tokens. |
| Cached input price | $0.01 per million cached input tokens. |
| Output-token charges | None. |
| Questions per request | Up to 400, twice the previous limit. |
Customer tests show gains, with different comparisons
Box evaluated document routing on 30 certificate-of-insurance cases. Biswaroop Palit, its vice president of product management for platform and integrations, said route selection was about five times faster than with the Responses API, with the same accuracy. Lower-confidence cases were sent to reviewers.
Datadog is using the service to assess the risk of agent-generated rshell commands before execution. Staff engineer AJ Stuyvenberg said it delivered 40 times faster p95 latency than available frontier models, with comparable accuracy. The p95 measure describes the response time within which 95% of requests finish.
These results cover different tasks and comparison systems. They demonstrate specific customer uses but do not establish that every deployment will achieve the same speed improvement or accuracy.
Pricing and deployment options shape implementation
OpenAI charges for input tokens without a separate output-token fee. Prompt caching provides a lower rate for reused context, making the pricing structure relevant to applications that repeatedly apply the same instructions.
Since the public beta, OpenAI has also added native GPT-Live integration for decisions during live conversations. Potential uses include deciding whether a support agent should ask a clarifying question, route a request or escalate an issue.
Regional processing is available in the US and EU, while Zero Data Retention is available to eligible customers. These are deployment options, not a statement that every customer automatically receives the same data controls.
Conclusion
The Decisions API is now available through a dedicated endpoint. The next practical step for developers is testing a recurring decision against representative examples in the Decisions Playground, checking whether the speed, accuracy and cost fit their own workflow.
Read More News
Disclaimer
This FinanceVelix article provides general information, not personalized financial, investment, legal, or tax advice. Market prices, interest rates and economic conditions can change; investments carry a risk of loss. Facts reflect the material available when published and may change. Check important information independently and consult a qualified professional when needed. Mentioning an asset, company, or service is not an endorsement.
