What happened
On October 6, 2026, OpenAI released its Decisions API in public beta for all developers. Announced on OpenAI’s developer community, the new endpoint is designed to let apps “choose the right model, tool, or action in near real-time” — returning a probability, a fixed choice, or a numeric score instead of prose. OpenAI claims decisions come back up to 10x faster than the same model queried through the Responses API.
The details
The Decisions API runs on GPT-6 Luna — OpenAI’s low-cost reasoning model, released September 22 alongside GPT-6 Sol — and accepts text and image inputs. It’s a separate route (POST /v1/decisions) with three question types:
- Predicates: estimate the probability that a statement is true.
- Choices: pick from a predefined list of options, with per-option probabilities and a confidence score.
- Scores: evaluate an input against a numeric range.
Because the output space is closed, the model can’t answer outside the list — there’s no free text to parse, and any question can return a refusal rather than a bad answer. DevDay slides put the speed at roughly 150 milliseconds against about 1.6 seconds for a standard Luna call (vendor-stated; no independent benchmark yet). The API launched at DevDay 2026 on September 29 as a limited preview, with broad availability promised in the days that followed — and GA expected “in the coming weeks.” OpenAI’s API guide lists pricing at $0.10 per million input tokens, with no output or cache charges.
The release also lands in the middle of a growing “decision model” category: Diogo Almeida’s TypeSafe AI released Jev in mid-September, Cloudflare shipped Clef on October 2, and AWS open-sourced Strands Decider 2B — all betting that developers want fast typed answers, not text, to drive agents and routing logic.
Why it matters
Every agentic app has a step where the LLM stops writing and starts deciding — route this request, classify that content, pick the agent’s next action. Paying for full text generation to get one word back is slow and wasteful. A typed, 150-millisecond endpoint is exactly the “system one” fast-thinking layer that agentic architectures need, and OpenAI making it available to all developers (with a Playground for experimenting before writing code) lowers the barrier for everyone building agents on its platform. Keep the honest framing in mind: the 10x figure is OpenAI’s own claim, published without a benchmark, and real latency depends on your region, input size, and traffic.
Bottom line
OpenAI’s Decisions API is now in public beta: a fast, cheap endpoint powered by GPT-6 Luna that returns typed decisions (probability, choice, or score) in about 150ms — 10x faster than the standard API, per OpenAI. Useful anywhere an agent needs to decide, not write.
Frequently asked questions
What is the Decisions API?
A new OpenAI API endpoint (POST /v1/decisions) that returns a typed answer — a probability, a fixed choice, or a numeric score — based on text or image context, instead of generating free text.
How fast is it, really?
OpenAI claims decisions are up to 10x faster than GPT-6 Luna through the Responses API — roughly 150ms versus 1.6s, based on DevDay slides. No independent benchmark has been published yet.
How much does it cost?
OpenAI’s API guide lists $0.10 per million input tokens, with no output or cache charges. Regional processing premiums and long-context multipliers apply.
When will it be generally available?
OpenAI says GA is expected in the coming weeks. It’s in public beta for all developers as of October 6, 2026, and can be tested in OpenAI’s Playground.
Sources: OpenAI Developer Community announcement, Unite.AI, The Decoder