OpenAI Decisions API: Pricing, Routing and Limits

OpenAI made its Decisions API generally available on October 22, 2026. Rather than asking a general-purpose assistant to write a paragraph, developers can ask a narrowly defined question about text, an image or structured data and receive a typed answer. The release targets an expensive but easily overlooked part of automation: small classification, scoring and routing decisions repeated thousands of times.

What is now available

The dedicated POST /v1/decisions endpoint currently supports gpt-6-luna. Each request supplies shared input and between one and 400 questions. The three supported question types are predicate, which estimates the probability of a condition; choice, which selects from predefined options; and score, which rates an input against ordered levels. OpenAI also added native JSON input, images in question instructions, prompt caching and integration with GPT-Live. These are bounded decisions, not a substitute for free-form generation or a full agent planning loop.

Where a bounded decision saves work

Consider a sales inbox where each incoming message needs a department, urgency score and whether it contains a specific purchase request. One Decisions call can assess those independent questions against the same message. The application can then route high-confidence messages automatically and put uncertain or sensitive ones into a human-review queue. The API returns probabilities and confidence signals; it does not independently establish that a decision is correct. Teams must evaluate results against labeled examples before defining production thresholds.

Latency and pricing need careful interpretation

OpenAI describes the API as up to ten times faster than using GPT-6 Luna through Responses for relevant bounded tasks, and cites customer-specific speed improvements. Those are not universal service-level guarantees. Actual latency depends on input size, region, workload and integration. At the published standard rate, one million input tokens costs $0.10, while eligible cached reads cost $0.01; there is no output-token charge. Cache reuse has matching and minimum-prefix conditions, and cache-write pricing may differ. Test a realistic workload rather than multiplying the headline speedup into a guaranteed business result.

A safe first implementation

1. Select a repeated binary, categorical or ordered-severity decision with a known correct outcome.
2. Create a labeled sample set including ambiguous, adversarial and high-impact edge cases.
3. Define distinct options or score levels; keep questions about observable evidence.
4. Compare Decisions against the existing process on accuracy, latency, cost and human corrections.
5. Set thresholds for manual review and monitor drift before enabling automatic downstream actions.

The endpoint can answer up to 400 independent questions about shared input, but sequentially dependent decisions require separate requests. This matters for workflows where a second question only makes sense after the first is resolved. OpenAI documents US and EU inference residency options and Zero Data Retention for eligible customers; eligibility and data handling should be checked against the current official controls before sending sensitive information.