OpenAI Decisions API Generally Available, Powered by GPT-6 Luna

What happened

Before going GA, the API went through two stages: a limited preview at the Sept 29 developer day and a public beta on Oct 6, reaching general availability about three weeks later. Requests go to the dedicated endpoint POST /v1/decisions, and per OpenAI's developer docs, gpt-6-luna is currently the only supported model.

The API targets high-frequency, low-latency decision calls: a single request can carry 1 to 400 questions, twice the beta limit, with text, image, and native JSON inputs. OpenAI says its own teams already use it: the dots team to cut response latency, the personal finance team to categorize transaction descriptions, and the communications team to classify coverage by theme and score it.

Between beta and GA, OpenAI added native GPT-Live integration (triggering decisions mid-voice-session), prompt caching, inference residency in the US and EU, plus usage reporting and Private Link support; Zero Data Retention and HIPAA compliance are available to eligible customers.

Key facts

1. Three question types: predicate returns a 0–1 probability; choice returns one preset option with per-option probabilities and a confidence value; score grades on a 2–10 level ordered rubric and returns the probability-weighted average
2. Pricing: $0.10 per million input tokens, $0.01 per million cached input tokens, no output-token charge
3. Speed: the company says about 10x faster than calling GPT-6 Luna via the Responses API
4. Customer evaluations (quoted in the announcement): Box reports about 5x faster route selection (30 cases, same accuracy); Datadog reports 40x faster p95 latency scoring agent-command risk; Sierra reports p95 under 280ms at 90% lower cost per call
5. Around the same time: GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna entered private preview in Google Cloud's Model Garden on Oct 21