OpenAI's Decisions API Is Generally Available: Typed Answers, $0.10 per Million Input Tokens and What to Test

OpenAI made its Decisions API generally available on October 22. It is a dedicated endpoint for small questions with bounded answers, such as classify this, route that, or should the agent retry. OpenAI says it is up to 10 times faster than calling GPT-6 Luna through the Responses API, and that you pay only for input tokens. The speed figures are OpenAI's and its customers' own claims, and I have not run the API.

I read the launch post, Build more responsive apps with the Decisions API, and the start of the developer guide, Decisions. I did not read the pricing guide, the API reference or the data controls pages that the post links.

What it is

Per the post, you define a question and its allowed answers, then give context as text, images or native JSON. The API returns one of three typed answers:

Type	Use	Result
predicate	Estimate whether a condition is true	A probability from 0 to 1
choice	Pick one of your options	One of your values, plus probabilities and confidence
score	Rate against ordered rubric levels	The probability-weighted average of the level indices

The guide says requests go to POST /v1/decisions using gpt-6-luna, which it calls the only currently supported model.

What changed since the public beta

The post lists: a native GPT-Live integration, prompt caching for reused context, inference residency in the US and EU, native JSON inputs, images in question instructions, up to 400 questions per request (twice the previous limit), and added capacity in several Asian regions to improve latency. It says Zero Data Retention is available to eligible customers.

Price

OpenAI states $0.10 per million input tokens and $0.01 per million for cached input, with no output-token charges. It points to a pricing guide for cached-write details and says to check current rates there. I did not read that guide, so check it before you plan around the numbers.

Customer claims in the post

These are quotes chosen by OpenAI, not independent tests:

• Box says that on 30 certificate-of-insurance cases it saw about 5x faster route selection with the same accuracy as the Responses API. Thirty cases is a small sample.
• Datadog says it uses the API to score the risk of agent-generated shell commands before running them, with 40x faster p95 than available frontier models at comparable accuracy.
• Ramp says attributing AI costs was 4x faster and 30% cheaper than with GPT-6 Luna.
• Sierra says it runs under 280 ms p95, about 3x faster and 90% lower cost per call than its previous approach, with no drop in customer outcomes.
• Speak says the GPT-Live integration cut time to next action by over 90%, from 1.4 seconds to 0.13 seconds.

Each company describes its own setup. None of these numbers transfer to your workload automatically.

Why it is interesting for agents

Agents make many small decisions between the big model calls. A fast, cheap, typed answer fits there: choose the next action, decide whether to ask for clarification, or score a risky command. The Datadog example is the notable one, since it puts a model in the path of a remediation action. That is the customer's claim about its own use, and a wrong decision there has real cost.

What I would test

These are my own untested suggestions.

1. Replay your logs. Take real inputs where you already know the right label and compare Decisions against your current method on accuracy first.

2. Use the confidence outputs. The choice type returns probabilities and confidence. Set a threshold and send low-confidence cases to a person or a stronger model, which is the pattern Box describes.

3. Measure p95 and p99 latency from your region, not averages. Check the regions you use.

4. Calibrate. Check whether a stated 0.9 probability is right about 90% of the time on your data.

5. Do not gate dangerous actions on it alone. If you score command risk, keep hard policy checks next to it.

6. Price your real input sizes. Long inputs cost more even with no output charge. Compare with your current cost per call.

7. Check data terms. Look at the data controls page for residency, retention and ZDR eligibility.

What I would not conclude

• That it will match a larger model on open-ended tasks. It is built for bounded answers.
• That the 10x speedup holds for every question type or input size.
• That output being free means total cost is low. Input tokens still count.

Frequently asked questions

What is the Decisions API?

A dedicated OpenAI endpoint that returns a probability, a choice or a score for a bounded question, per the launch post.

Which model does it use?

GPT-6 Luna, the only supported model at the time of the guide I read.

How much does it cost?

$0.10 per million input tokens and $0.01 per million cached input, with no output-token charge, according to OpenAI.

How fast is it?

OpenAI says up to 10x faster than GPT-6 Luna through the Responses API. Customer figures in the post vary.

How many questions can one request hold?

Up to 400, per the post.

Does it work with voice?

Yes. The post describes a native GPT-Live integration that emits decision events from live sessions.

Is it available in the EU?

The post says inference residency is available in the US and EU. Check the data controls page for terms.

Sources

• OpenAI, Build more responsive apps with the Decisions API (October 22, 2026): launch claims, pricing and customer quotes.
• OpenAI developer docs, Decisions: endpoint, model and what's new. Only the opening of the guide was read.