OpenAI Decisions API Goes GA, Output Free

OpenAI on Thursday, Oct. 22, made its Decisions API generally available, opening to all developers an endpoint that returns probabilities, choices and scores instead of generated text. The service runs only on GPT-6 Luna and charges nothing for output tokens, according to OpenAI's developer documentation and reports by WebProNews and Seeking Alpha.

Key takeaways

• The Decisions API left public beta on Oct. 22 and now accepts up to 400 questions per request, double the beta limit of 200.
• Uncached input costs $0.10 per million tokens and cached input $0.01, with no charge for output tokens.
• OpenAI says the endpoint is up to 10 times faster than running the same model through the Responses API; early customers including Datadog, Sierra and Speak report their own latency gains.

What the OpenAI Decisions API returns

Developers send a request to POST /v1/decisions with an input — text, images or JSON — and a list of named questions. The model answers each question in one of three typed formats rather than writing prose.

A predicate question returns a probability from 0 to 1 that a condition is true. A choice question picks one of the values the developer supplies and attaches probabilities and a confidence figure. A score question rates input on two to 10 ordered levels and returns a probability-weighted average. The calling application decides what to do with each answer.

Decisions API pricing and limits

Pricing is input-only: $0.10 per million uncached input tokens and $0.01 per million cached tokens. Cache writes are billed, and regional processing premiums and long-context multipliers still apply. Caching kicks in only when the shared input prefix reaches 1,024 tokens.

Since the beta, OpenAI has added prompt caching, inference residency in the United States and Europe, and native JSON and image input, according to WebProNews. Zero Data Retention and HIPAA use are available to eligible customers.

Customer results on latency and cost

The figures below are self-reported by customers, as relayed in launch coverage, and have not been independently benchmarked.

Speak used the API with GPT-Live. Chief technology officer Andrew Hsu said the company "was able to reduce the time to next action by over 90%, from 1.4s to 0.13s." Datadog uses it to rate the risk of shell commands written by AI agents and reported 40 times faster p95 latency with accuracy comparable to frontier models.

Sierra, which picks next actions in live customer conversations, reported p95 latency under 280 milliseconds, about three times faster and 90% cheaper per call than its previous approach. Ramp said AI cost attribution became four times faster and 30% cheaper compared with GPT-6 Luna, and Box reported roughly five times faster route selection across 30 certificate-of-insurance cases.

Context: a race to own the agent control loop

The launch targets the small, frequent decisions inside AI agents — continue, retry, escalate or stop — that are often handled today by asking a large model to emit JSON.

OpenAI is not alone. TypeSafe AI released its decision model Jev on Sept. 15, and AI Newsway reported that the startup left stealth with $40 million on the same premise.

What's next

The API is limited to GPT-6 Luna for now. For outputs that need a custom JSON schema, OpenAI still points developers to Structured Outputs in the Responses API. The open question is whether independent tests confirm the speed claims at production scale, and whether OpenAI extends the endpoint to other models.

FAQ

Is the Decisions API a new model?

No. It is a new endpoint that serves GPT-6 Luna, the only supported model, and returns typed answers instead of text.

How much does it cost?

Uncached input is $0.10 per million tokens and cached input is $0.01 per million. Output tokens are free, though cache writes, regional premiums and long-context multipliers apply.

When should developers use Responses instead?

When they need generated text or a custom JSON schema. OpenAI recommends Structured Outputs in the Responses API for those cases.

References

• Decisions — OpenAI API documentation
• OpenAI Makes Its Decisions API Generally Available, Billing Only for Input Tokens — WebProNews (Oct. 22, 2026)
• OpenAI launches Decisions API with up to 10x faster AI decisions — Seeking Alpha via TradingView (2026)