OpenAI Decisions API Review: Typed Verdicts, About 10x Faster Classification

On October 6, 2026, OpenAI put the Decisions API into public beta (a limited preview appeared at DevDay on September 29). The endpoint is POST /v1/decisions and currently runs a single model, gpt-6-luna (1,050,000-token context). Three question types: predicate returns a 0–1 probability that a condition is true; choice picks from fixed options with confidence; score rates against ordered severity levels. Image input must be inline base64 data URLs; hosted links and file IDs are refused. Zero Data Retention and HIPAA eligibility are supported, with data residency in the US and Europe. OpenAI says general availability is coming in the next few weeks.

The basis is the published spec and vendor claims (no published independent end-to-end benchmark yet): $0.10 per 1M input tokens, free output and cache reads/writes; OpenAI says it is about 10x faster than the Responses API (vendor claim, no published benchmark or compute disclosure). Public figures for comparison: TypeSafe Jev (early access, September 15) charges $0.042 per 1M input tokens with free output, reports 70–500 ms end-to-end measured from the US West Coast, and supports up to 255 choices. Note: neither vendor has published head-to-head accuracy numbers.

1. Cost contrast: The same gpt-6-luna costs $0.50 per 1M output tokens via the Responses API, $0 via Decisions.
2. Image limits: Inline base64 only; hosted image URLs and uploaded file IDs are refused.
3. Voice: Decisions can drive actions via client delegation with the Live API.