OpenAI just made it official: stop paying for prose you never needed. The Decisions API — in public beta since October 6 — returns typed verdicts instead of paragraphs: a probability, a choice, or a score. It runs on GPT-6 Luna, costs $0.10 per million input tokens with zero output charges, and OpenAI says it is about 10x faster than the Responses API.
Asked a model “which department handles this ticket?” and then parsed the answer out of a polite paragraph? This endpoint skips the essay and hands your code something actionable.
Three question types, zero small talk
Every request hits POST /v1/decisions with three parts: the model (gpt-6-luna, the only option for now), your input (text, images, or both), and your questions. Answers come back in the typed shape you defined.
| Type | Use it for | What you get back |
|---|---|---|
| predicate | Check a condition — visible damage? relevant passage? | probability from 0 to 1 that the condition is true |
| choice | Pick from fixed options — department, category | The chosen value, probabilities per option, plus a confidence score |
| score | Rate against ordered levels — issue severity | A probability-weighted score that can land between levels |
Source: OpenAI Decisions API documentation.
A predicate suits “flag this if you are 90% sure.” A choice shows its work — per-option probabilities plus a separate confidence value for your own thresholds. A score handles ordered rubrics like severity, returning a weighted average across levels.
The pricing math
Decisions API pricing: $0.10 per million input tokens, $0 for output tokens, $0 for cache operations. Third-party analysis estimates that cuts classification workloads to roughly one-fifth of standard Luna pricing.
| Decisions API | GPT-6 Luna via Responses API* | |
|---|---|---|
| Input | $0.10 / 1M tokens | ~$0.10 / 1M tokens |
| Output | $0 | ~$0.50 / 1M tokens |
| Cache read / write | $0 | Charged |
*Estimates based on third-party coverage of OpenAI’s pricing. Official Decisions API pricing from OpenAI’s docs.
Work the example: a 500-token support ticket costs about $0.00005 to classify. Run a million of them and you have spent roughly $50 — with no output-token surprises. That is the kind of unit economics that makes always-on moderation and routing genuinely affordable.
Stat highlights (per OpenAI)
- ~10x faster than the Responses API — OpenAI’s claim for typed answers vs generated prose
- Oct 6, 2026 — public beta launch; general availability expected “in the coming weeks”
- Text + images accepted as input; Zero Data Retention and HIPAA available for eligible customers
What it replaces
The old pattern: send a ticket to a chat model, get a paragraph, parse it with regex, pray the format never drifts. Slow, expensive, fragile. The Decisions API collapses that into one typed call.
From paragraphs to verdicts
Your text and/or images
predicate · choice · score
probability · choice · score
route · flag · prioritize
No prose in between. No parsing. No format drift.
How to profit: sell decision pipelines
Every company with a support inbox, moderation queue, or QA process pays humans — or pricey LLM calls — to make repetitive judgment calls. Productize it in a week.
- Find a client drowning in decisions. Support ticket routing, content moderation, product-photo QA, lead scoring — anywhere someone reads and sorts all day.
- Define 3–5 crisp questions with distinct choices. Always include an “other” fallback so edge cases hit a review queue, not an error.
- Prototype in the Playground. OpenAI’s docs link a no-code playground — test your questions on real client examples before writing a single line of code.
- Wire up POST /v1/decisions. Send the input once, ask several independent questions per request, and set routing thresholds from labeled examples.
- Keep a human in the loop for low-confidence answers. The confidence field tells you exactly when to escalate — clients love this because it bounds the risk.
- Charge for the pipeline, not the tokens. A $1,500/month “smart triage” retainer against roughly $50 in API costs per million decisions is the kind of margin that builds a business. (Illustrative.)
Watch the fine print
- It is a public beta — expect rough edges, and confirm behavior before betting a client SLA on it.
- Only gpt-6-luna is supported for now.
- Images must be inline base64 — hosted image URLs and uploaded file IDs are refused.
- Probabilities are the model’s estimates — calibrate thresholds on your own labeled data.
- The 10x speed figure is OpenAI’s claim, measured against its own Responses API.
Keep Reading
Frequently asked questions
Is the Decisions API generally available?
Not yet. It launched in public beta on October 6, 2026, and OpenAI expects general availability “in the coming weeks.”
Which models does it support?
Only gpt-6-luna at launch — a reasoning model OpenAI released alongside GPT-6 Sol in September 2026.
Why is it faster than the Responses API?
It skips text generation and returns typed answers — about 10x faster, per OpenAI. No tokens spent on sentences nobody reads.
Can it process images?
Yes — text, images, or both. But images must be inline base64 data URLs; hosted HTTP/HTTPS URLs and file IDs are not supported.
How is this different from JSON mode?
Structured Outputs still generates prose shaped to your schema — and bills you for it. Decisions returns probabilities, choices, and scores with zero output charges.
Is my data private?
Zero Data Retention and HIPAA use are available for eligible customers, with data residency in the United States and Europe (EEA + Switzerland).
What will it cost at scale?
$0.10 per million input tokens, nothing for output or cache — about $50 per million 500-token classifications, per third-party estimates.