OpenAI's Decisions API is generally available, with 400 questions per request and prompt caching
OpenAI took the Decisions API out of beta on 22 October 2026 (blog; changelog; forum post, 17:51 BRT). It had been in public beta since 6 October.
What changed since the beta
- Up to 400 questions per request (was 200).
- Prompt caching: cached input costs $0.01 per 1M tokens; uncached input stays at $0.10. No output charges.
- Native JSON input (
input_json), images inside individual question instructions, hosted image URLs, and conversation history with tool calls. - GPT-Live integration, so voice apps can call decisions while people speak.
- US and EU inference residency, Private Link and usage reporting.
The model is still gpt-6-luna, the only one the endpoint supports. OpenAI says it is “up to 10x faster” than the same model through the Responses API and quotes customers Sierra and Speak (UnverifiedClaim). TechCrunch frames it as OpenAI’s answer to Jev.
On the boards
Independent numbers are below Jev: JevBench API board Intelligence 56.9 (Jev 1.13: 63.6), and 48.5 for “GPT-6 Luna (Decisions)” on the Artificial Analysis Decision Index (Jev 1.14: 57).
Still a runtime
GA did not add decision-trained weights or a decision head, so the runtime page stays in (NOT)Models and off the quadrant.