OpenAI’s Decisions API exits beta at 400 questions a call

OpenAI took its Decisions API out of public beta on October 22 and raised the limit to 400 questions per request, up from 200 during the beta. In its announcement, OpenAI says the GPT-6 Luna-based endpoint returns a probability, a choice or a score up to 10x faster than Luna through the Responses API. If your team sends Luna lots of small classification or routing calls, most of the savings come from how Decisions bills: input costs the same $0.10 per million tokens as Luna anywhere else, but output is free and one shared context can carry up to 400 questions.
That gap is large. We ran a workload of 400 decisions over shared context at list prices: 400 separate Responses calls, each resending the 2,050-token context and producing five output tokens, come to $0.083, while a single Decisions request comes to $0.0022. Luna’s output on the Responses API costs $0.50 per million tokens, and that charge doesn’t exist here. Cached input is $0.01 per million.
What changed since the beta
The beta-to-GA list adds prompt caching for reused context, native JSON inputs, images inside question instructions, inference residency in the US and EU, and a native GPT-Live integration. OpenAI also says it added capacity in several regions in Asia to cut latency. The basic contract is the same: you define a question and its allowed answers, pass context as text, images or JSON, and get back one of three answer types:
predicate: a probability from 0 to 1 that a condition is true, such as whether a product shows visible damage.choice: one of your predefined values, plus probabilities and confidence, such as whether an agent should continue, retry or ask for clarification.score: the probability-weighted average of level indices on an ordered rubric, such as damage severity.
The 10x figure is OpenAI’s own and compares Decisions with the same model on the Responses API. When the API first showed up as a limited preview at DevDay in September, OpenAI said it answered in 150 milliseconds, against 1.6 seconds for Luna.
To use it with GPT-Live, a voice app adds a Decisions configuration to session.triggers. The session emits a decision.event, and the app acts on the selected answer. OpenAI’s examples are a support agent deciding whether to ask a clarifying question, route a request or escalate it, and an avatar that reacts to a user’s words with a wink or a smile. Inside OpenAI, the Personal Finance team uses Decisions to categorize transaction descriptions. The Communications team uses it to sort press coverage by theme and to score whether each piece reinforces or challenges key narratives.
Our analysis
What we checked
On October 23 we read OpenAI’s Decisions developer guide, API reference, pricing page and prompt-caching guide, compared them with archived beta pages from October 11 and 12, and priced sample workloads at list rates. We did not call the API, so all of this comes from the documentation, not from measurements.
OpenAI’s post points readers to rates for 30-minute and 12-hour cached writes. The Decisions API reference, though, says 30 minutes is currently the only cache lifetime this endpoint supports, and the 12-hour rate is listed for the Responses API. On a request with 400 questions, caching cuts the bill by only 8%, because question text is never cached.
| 2,000-token shared context + 400 questions of 50 tokens | Per request | Per 1,000 requests |
|---|---|---|
| Decisions, uncached | $0.0022 | $2.20 |
| Decisions, context cached (steady state, write excluded) | $0.00202 | $2.02 |
| Same tokens at Microsoft-Decision-1’s listed $0.042 per million | $0.000924 | $0.924 |
- A 30-minute cache write on Decisions costs $0.125 per million tokens. It is charged instead of the input rate, not on top of it. With reads at $0.01 and uncached input at $0.10, a cached prefix is cheaper from its second send. The prefix has to be at least 1,024 tokens, and question content, including question-specific images, doesn’t count toward that.
- At list rates, Decisions costs 2.38 times Microsoft’s per-token price uncached and 2.19 times with the context cached. Token counts differ between tokenizers and request formats, so these figures compare rates and are not measured bills.
- The guide says regional processing premiums apply, and the pricing page sets a 10% uplift for models released on or after March 5, 2026. By our inference, Luna falls under that rule, which would make residency input $0.11 per million.
How to get started
Decisions is available now with GPT-6 Luna through its own dedicated endpoint. Regional processing runs in the US and EU, and Zero Data Retention is open to eligible customers. OpenAI suggests starting with one decision your application makes often: define the question, the allowed answers and the context needed to tell them apart. Then test it on examples that reflect real usage, either in the Decisions Playground with your own text or images, or by following the developer guide to a first request.
Up next
Claude apps switch to auto mode by default next week
Claude won’t ask you to approve each step by default starting next week, because an automated check will block risky actions on Pro, Max and Team plans instead. Will you keep auto mode on in Claude or switch back to manual?
Related stories
- OpenAI’s Decisions API beats its own engine by up to 10x
- OpenAI’s Decisions API picks an answer in 150 milliseconds
- Gemini Decisions API scores 1,000 questions in one call
- OpenAI keeps Codex cache forever, API cache gets 12 hours
- Your ChatGPT Plus plan can now pay for other apps’ AI
- Code strings point to a $50 production plan for OpenAI’s API
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

