Decisions API is now GA: cache shared evidence, pass JSON directly, and let GPT-Live call it

#What changed

OpenAI Developers announced on Oct 22, 2026 that the Decisions API, which returns typed yes/no probabilities, picks from a fixed set, or rubric scores without generating a full response, is now generally available. Since the public beta (covered here on Oct 6), the docs list these additions: prompt caching on shared input (on by default, needs a prefix of at least 1,024 tokens), images inside individual question instructions, native JSON input via an input_json content part, up to 400 questions per request (up from 200), hosted image URLs as well as base64, conversation history with tool calls and outputs, organization usage reporting at GET /v1/organization/usage/decisions, a native GPT-Live integration, and Private Link. It still only runs on gpt-6-luna, at $0.10 per 1M uncached input tokens with no charge for output tokens.

#Who is affected

Developers calling POST /v1/decisions for routing, classification, moderation, or review gates, and teams building voice apps on GPT-Live. This is an API change; it does not change ChatGPT, Codex, or Dots plans or usage limits.

#What to do now

• If you send the same long reference (a policy doc, a product image) with different questions, keep that shared input byte-identical across calls so the cache hits, and check usage.input_tokens_details.cached_tokens to confirm it.
• Stop stringifying structured records: pass them as an input_json part in a user message. Keys inside it are treated as plain data.
• If you were splitting batches at 200 questions, you can now go to 400 per request, but keep dependent questions in separate calls.
• Set thresholds from your own labeled examples, and log each decision next to its input before letting it trigger anything that spends money, since a wrong route fails silently.

#What is not confirmed

OpenAI hasn't published calibration data for the returned probabilities or latency numbers against a small self-trained classifier in the GA post; several developers asked about both in the replies. The "about 10x faster than the Responses API" figure is OpenAI's own claim, not an independent benchmark.