OpenAI announced general availability for its Decisions API on October 22, following an October 6 public beta. The company presents it as a specialized interface for bounded answers—a probability, a choice, or a score. The production question is whether those answers deserve authority inside a particular workflow. OpenAI’s announcement establishes the release, but does not settle that question.
What a typed answer actually means
The Decisions guide describes three answer types. A predicate estimates a condition’s probability. A choice selects a supplied option and includes probabilities and confidence. A score is the probability-weighted average of ordered level indices, so it can fall between levels. The application determines what happens afterward.
That separation matters. A number can be easy to consume without being sufficient evidence for an action. Developers should distinguish an answer that fits the interface from one that meets their operational standard. The practical question is what evidence would justify letting that answer change a case’s destination or outcome.
What general availability adds
The October 22 changelog confirms shared-input prompt caching, native JSON input, images within individual question instructions, and an increase from 200 to 400 questions per request. OpenAI’s announcement also lists native GPT-Live integration.
As checked October 24, the API reference identifies gpt-6-luna as the only supported model. Requests go to POST /v1/decisions; structured JSON is limited to 32 container levels, 10,000 values, and 256 KiB of compact UTF-8 JSON.
For voice applications, this adds a decision-layer question to the interaction challenges discussed in our earlier GPT-Live-1 coverage: which judgments should control the next step? Integration alone cannot answer that design question.
The speed claim has a narrow evidentiary base
OpenAI claims decisions can be up to 10 times faster than GPT-6 Luna through the Responses API. Its announcement includes a Box comparison involving 30 certificate-of-insurance cases. Those are vendor and selected customer accounts, not independent benchmark findings. The reviewed materials do not provide a complete methodology establishing how the headline improvement transfers across workloads. The announcement should be read with that limitation.
Thresholds need evidence from the intended application
OpenAI’s guide recommends labeled application examples and thresholds that reflect false-positive and false-negative costs. That makes evaluation central to adoption, rather than a final check after wiring an answer into an action. The guidance does not establish that any particular threshold will work.
A useful evaluation would compare representative cases with the existing workflow, examine consequential mistakes, and define a route for uncertain cases. Human review is one editorial recommendation, not a reported endpoint feature. Teams should also decide what outcome would cause them to revise a threshold or withdraw automation.
No API testing was performed for this report. The available evidence establishes the interface and OpenAI’s claims; it does not independently establish accuracy, probability calibration, or production reliability.
Cost and fit remain workload decisions
The guide lists uncached input at $0.10 per million tokens with no output-token charge. Cache writes cost extra; regional and long-context multipliers apply. It directs developers toward Responses for custom JSON schemas or explanations, and function calling for tool requests.
The adoption boundary is therefore practical: does the task fit a bounded answer, and does evaluation justify the action attached to it? More questions per request do not themselves establish lower cost per decision. General availability opens an implementation option; a team’s own evidence must establish its value.
Sources
- OpenAI announcement, October 22, 2026.
- Decisions guide, checked October 24, 2026 KST.
- Create a decision reference, checked October 24, 2026 KST.
- OpenAI API changelog, October 22, 2026 entry.