OpenAI Decisions API Generally Available GPT-6 Luna Price: $0.10 Per Million Tokens, 10x Faster Typed Answers Explained
Image: Unsplash (free Unsplash licence) — Illustrative image: an analytics dashboard of the kind the Decisions API scores and routes
📌 KEY HIGHLIGHTS
- Generally available from 22 October 2026: the Decisions API is now a production endpoint, POST /v1/decisions, after a DevDay preview on 29 September and a public beta from 6 October.
- One model only: gpt-6-luna is the sole supported model. You send shared evidence — text, images or JSON — plus named questions, and get back an answers array your code can act on directly. No paragraph to parse.
- Three answer types: predicate (a probability from 0 to 1), choice (one of your own options, with probabilities and a confidence value) and score (a rating against 2–10 ordered rubric levels).
- What GA added: up to 400 questions per request (doubled from 200 in beta), native JSON input, hosted image URLs as well as base64, images inside question instructions, prompt caching, conversation history with tool calls, organisation usage reporting, GPT-Live integration and Private Link.
- The price: $0.10 per million input tokens, $0.01 per million cached input tokens, and no output-token charge at all — because the API returns typed answers, not generated text.
What happened?
Most AI APIs answer in words: you ask a question, the model writes a reply, and your code then has to dig the actual decision out of that text. The Decisions API removes that step. OpenAI built it for the huge number of calls inside real applications that are not conversations at all — is this photo damaged, which queue should this ticket go to, how severe is this bug, should this request be escalated? For those, a paragraph is waste. What the application needs is a number or a label it can branch on, delivered fast and cheaply.
That is exactly what general availability formalises. The endpoint evaluates text, images and structured JSON against up to 400 developer-defined questions in a single request, and OpenAI says it returns those typed answers about 10 times faster than asking the same GPT-6 Luna model through the Responses API. Treat that 10x figure as OpenAI's own claim for now — the company has not published a benchmark, and no independent latency test had appeared by publication time. The direction, though, is easy to believe: the API skips generating output tokens entirely, which is where most of the time in a normal model call goes.
OpenAI Decisions API generally available GPT-6 Luna price — the three question types
| Question type | What you ask | What comes back |
|---|---|---|
| Predicate | Is a condition true? (e.g. does this product photo show visible damage?) | A probability from 0 to 1 that the condition is true — your code sets the threshold |
| Choice | Pick one option (e.g. billing, technical, shipping or other for a complaint) | The chosen value, a probability for every option, and a confidence value |
| Score | Rate against a rubric (e.g. bug severity across 2–10 ordered levels) | The probability-weighted average of the level indices — a score that can fall between levels |
| Price | $0.10 per million input tokens; $0.01 per million cached input tokens; no output charge. Regional processing premiums and long-context multipliers still apply | |
| Limits | gpt-6-luna only; 1–400 questions per request; files, audio and file_id inputs are not supported; decisions that depend on an earlier answer need separate requests | |
The GA release is also where the API grew up operationally. Prompt caching reuses a shared input prefix of at least 1,024 tokens at the discounted cache rate. A native GPT-Live integration lets a voice application run decisions as transcripts arrive — OpenAI's example picks an avatar's facial expression from the conversation, and a support agent can use the same pattern to decide whether to ask a clarifying question, route the request or escalate it. Inference residency is offered in the United States and the EU with added capacity in several Asia regions, Zero Data Retention and HIPAA use are available to eligible customers, and a Decisions Playground lets teams test questions against their own text or images before writing code. For the wider GPT-6 picture, see our explainer on OpenAI GPT-6 models arriving on Google Cloud Model Garden.
Image: Unsplash (free Unsplash licence) — Illustrative image: code on a screen
Why it matters for developers in India and worldwide
Classification, routing and scoring calls are the invisible bulk of AI spending: every support ticket triaged, every upload moderated, every transaction categorised is one of these small decisions, often made millions of times a day. At $0.10 per million input tokens with output free, a 500-token support message costs roughly $0.00005 to classify — cheap enough that the economics of "add a human review queue only for low-confidence cases" start to work for small teams, not just platforms. The customer results OpenAI quoted at launch are company-reported, not independently verified, but they show the shape of the gains: Box reported roughly 5x faster route selection across 30 certificate-of-insurance cases at the same accuracy as the Responses API, Datadog reported 40x faster p95 latency when scoring the risk of agent-generated shell commands, Ramp reported cost attribution 4x faster and 30% cheaper, Sierra reported under 280 milliseconds p95 in live conversations at 90% lower cost per call, and Speak reported cutting time-to-next-action from 1.4 seconds to 0.13 seconds with the GPT-Live integration. This is also the pattern proactive agents depend on: our Claude Duet proactive agent explainer shows the other side of the same shift — agents that act between messages need fast, bounded decisions underneath them — and our NVIDIA Nemotron 3.5 open models explainer covers the open-model route to cheaper agent plumbing.
Should you care?
✅ Care — and test it this week — if your app already pays a full model to answer questions you then parse: ticket routing, content classification, photo checks, severity scoring and agent next-step selection are the sweet spot, and the Playground makes a same-day evaluation realistic. ⏳ Wait, or stay on the Responses API, if: you need generated text, extracted fields in your own JSON schema or tool calls with arguments — OpenAI itself points those jobs to Structured Outputs and function calling — or if a single supported model and vendor-supplied speed claims are not enough for your risk bar. The professional move is the one OpenAI's own documentation recommends: build a small labelled test set from your real traffic, set your thresholds from the cost of false positives versus false negatives, and measure the latency yourself before migrating anything that matters.
FAQs
What is the OpenAI Decisions API?
It is a dedicated OpenAI endpoint, POST /v1/decisions, that evaluates text, images or JSON and returns typed answers instead of generated text: a probability for a predicate question, one option for a choice question, or a rubric score for a score question. It is designed for classifying content, routing requests and prioritising work inside applications, and it runs only on the gpt-6-luna model.
How much does the OpenAI Decisions API cost?
It costs $0.10 per million input tokens and $0.01 per million cached input tokens, with no charge for output tokens. Regional processing premiums and long-context input multipliers can still apply. By comparison, using GPT-6 Luna normally costs $0.10 per million input tokens plus $0.50 per million output tokens, which the Decisions API avoids entirely because it generates no text.
Is the OpenAI Decisions API faster than the Responses API?
OpenAI says it returns typed answers about 10 times faster than asking GPT-6 Luna through the Responses API, because it skips output generation. That is the company's own figure — no independent benchmark had been published at the time of writing — and customers quoted at launch reported p95 latencies as low as under 280 milliseconds in live use. Developers should measure it on their own workloads.
Can the OpenAI Decisions API read images?
Yes. Images can be supplied as inline base64 data URLs or as publicly accessible HTTP or HTTPS URLs, either in the shared input or inside an individual question's instructions, so each question can compare its own image against shared evidence. File uploads via file_id, audio and item references are not supported.
TAKEAWAY
The Decisions API is OpenAI admitting that most production AI calls are judgements, not essays — and pricing them like it: a probability, a choice or a score, up to 400 at a time, for $0.10 per million input tokens.
Follow USA Tech on TechByCountry for the next hourly update.
Sources: OpenAI API documentation — Decisions guide (primary, "now generally available") • Unite.AI — OpenAI Makes Decisions API Generally Available with GPT-6 Luna (22 Oct 2026) • CryptoBriefing — OpenAI makes its Decisions API generally available at $0.10 per million input tokens (22 Oct 2026) • CellCog — OpenAI Decisions API: Price, Question Types, vs Jev (Oct 2026). Checked 23 Oct 2026.
Comments
Post a Comment