OpenAI Decisions API goes GA with gpt-6-luna: bounded answers, up to 10x faster than Responses

Why it matters

【Key points】

OpenAI on Oct 22, 2026 made the Decisions API generally available with gpt-6-luna. It answers bounded questions—predicate, choice, or score—from text, images, and structured inputs on POST /v1/decisions. OpenAI says it is up to 10x faster than calling GPT-6 Luna through the Responses API. This is a cheap decision API for agent hot paths, not a new frontier model release.

【Verifiable】

Primary: https://openai.com/index/decisions-api/. Changelog (https://developers.openai.com/api/docs/changelog, Oct 22): GA with gpt-6-luna; prompt caching; images in question instructions; input_json; up to 400 questions per request (was 200 in beta). Docs (https://developers.openai.com/api/docs/guides/decisions): predicate (probability), choice (option + probabilities/confidence), score (probability-weighted level average). Since beta: native GPT-Live integration; US and EU inference residency. Official pricing: $0.10/1M input tokens; cached input $0.01/1M; no output-token charges. Same-page customer metrics: Box ~5x faster route selection on 30 COI cases vs Responses (same accuracy); Datadog 40x faster p95 vs available frontier models; Ramp 4x faster and 30% cheaper AI-cost attribution vs GPT-6 Luna; Sierra under 280 ms p95, ~3× faster and 90% lower cost/call; Speak GPT-Live cut time-to-next-action over 90% (1.4s→0.13s).

【Our read】

1. GA of the Decisions API we covered at public beta—not Baseten's open decision models (mammoth-1 / elephant-1), and not a Luna capability leap. The bet is structured answers for routing, scoring, and live voice side channels.

2. The 10x claim is vs Responses on bounded questions, not vs every model. Customer quotes (Box, Datadog, Ramp, Sierra, Speak) are vendor-reported; order-of-magnitude signals, not independent benchmarks.

3. Watch: whether agents put Decisions on the default hot path; whether 400-question and cache discounts show up in real bills; whether open decision-model vendors compete on the same predicate/choice/score shape.