OpenAI Decisions API price starts at $0.10 per million uncached tokens on GPT-6 Luna

The OpenAI Decisions API price stands at $0.10 per million uncached input tokens when using GPT-6 Luna. Cached input tokens cost $0.01 per million. The endpoint launched in general availability on October 22, 2026, after a two-week beta period.

Cache writes follow standard GPT-6 Luna rates of $0.125 per million for a 30-minute window or $0.25 per million for a 12-hour window. Requests exceeding 272,000 input tokens incur a 2x multiplier on all input and cache charges. Regional processing in the US or EU adds a 10 percent premium.

OpenAI Decisions API price details and caching mechanics

Input tokens qualify for caching only when a shared prompt prefix reaches at least 1,024 tokens. Output tokens carry no separate charge. The API accepts text, base64 images or public URLs (maximum 128 images per request), and native JSON payloads. File uploads, audio, and function-calling tools remain unsupported.

Developers report typical per-request costs of $0.00004 for a 400-question batch at average prefix lengths. This pricing favors high-volume micro-decisions such as content classification or next-action scoring inside agent loops. Long-context workloads above 272K tokens double the effective rate, so teams often split prompts to stay under the threshold.

Organization tier determines rate limits. Build, Launch, and Grow plans each publish separate RPM and TPM caps for GPT-6 Luna. Usage appears in the organization endpoint but carries no Decisions-specific override. Teams should test against their current tier before scaling production traffic.

GPT-6 Luna latency limits and bounded question performance

OpenAI states the Decisions endpoint runs up to ten times faster than the same model called through the Responses API. Early adopters record median latencies of 100–180 ms and p95 values below 280 ms on Sierra and Speak workloads. Independent public benchmarks remain limited.

Each request supports up to 400 questions, an increase from the beta limit of 200. Supported answer formats include predicate probabilities, single-choice selections with probability distributions, and score outputs weighted by rubric levels. The 1.05-million-token context window allows long shared prefixes that improve cache hit rates.

Zero Data Retention and HIPAA options exist for qualifying accounts. Inference residency remains restricted to US and EU regions for now. Asia capacity expansions are expected but not yet confirmed.

Bounded decision use cases and comparison with alternatives

Common production patterns include routing support tickets, scoring loan risk attributes, selecting next agent actions, and classifying user intent in voice flows via GPT-Live integration. Box, Datadog, Ramp, and internal OpenAI teams have cited reduced cost attribution overhead and faster interactive loops compared with full Responses calls.

Teams evaluating alternatives often benchmark against TypeSafe Jev, which lists lower input pricing but restricts output to text only and lacks native image handling. Direct comparison on identical workloads shows the Decisions API ahead on latency and image support while remaining competitive on price once cache hits are factored in.

Before rollout, validate cache hit rates on representative prefixes and measure p95 latency under expected concurrency. Monitor the OpenAI changelog for any future adjustments to regional premiums or context multipliers.