OpenAI takes the Decisions API out of beta on GPT-6 Luna, and does not charge for output tokens
The developer forum post is timestamped 8:51 p.m. UTC on Thursday. The endpoint is POST /v1/decisions, and gpt-6-luna is the only model. A request can carry 400 questions, up from 200 in beta. Uncached input is $0.10 per million tokens. OpenAI’s guide says output tokens are not billed, which is a different line from Luna’s $0.50 output price on the standard sheet.
Feature4 minIn-depth
James Okonkwo4 min read
- OpenAI
- GPT-6 Luna
- Decisions API
- Agents
- Pricing

Photo: OpenAISource page
OpenAI made the Decisions API generally available on Thursday, on a single model, GPT-6 Luna. The company’s developer forum post is timestamped 8:51 p.m. UTC. The changelog for 22 October says the endpoint is generally available with gpt-6-luna, that a request can now carry up to 400 questions, up from 200 in the public beta, and that callers can cache shared input, pass structured data with input_json, and attach images to individual question instructions. The product post says the API is for questions with bounded answers, and that it is up to 10 times faster than calling GPT-6 Luna through the Responses API. That multiple is OpenAI’s comparison of two of its own endpoints. The post does not publish the test.
The guide sends requests to POST /v1/decisions and says gpt-6-luna is the only model the endpoint supports. A developer sets a question and the answers that are allowed, then passes context as text, images or native JSON. The API returns one of three types. A predicate is a probability, from 0 to 1, that a condition is true. A choice is one of the values the caller supplied, plus probabilities and a confidence field. A score is a probability-weighted average of the indices in an ordered rubric. The guide says the API returns the typed answer and the application runs whatever action that answer selects. Files, audio and item references are not accepted as input.
The customers in the post are quoting their own tests
The launch post’s examples are internal and then commercial, and the commercial numbers are the customers’ words as OpenAI prints them. Inside OpenAI, the post says the dots team is using the API to cut response times, the personal-finance team is categorizing transaction descriptions, and the communications team is classifying coverage by theme and by whether it supports or challenges a narrative. Biswaroop Palit, a vice president at Box, is quoted saying that on 30 certificates of insurance, Decisions picked a route about five times faster than the Responses API, at the same accuracy, and that lower-confidence cases still go to a reviewer. AJ Stuyvenberg, a staff engineer at Datadog, said the API scores the risk of commands an agent generates before those commands run, and that against available frontier models it was 40 times faster at the 95th percentile, with comparable accuracy.
In our testing, the Decisions API integration in GPT-Live was able to reduce the time to next action by over 90%, from 1.4s to 0.13s.
— Andrew Hsu, chief technology officer, Speak, quoted in OpenAI’s Decisions API post, Oct 22, 2026
Two more figures sit in the same post. Ajeet Grewal, an engineer at Sierra, said the API chooses the next action in parts of a live customer conversation, runs under 280 milliseconds at the 95th percentile, and is about three times faster and 90 percent cheaper per call than Sierra’s previous approach, with no drop in customer outcomes. René Sultan, a research engineer at Ramp, said attributing AI costs to the work that caused them is four times faster and 30 percent cheaper than using GPT-6 Luna, without a loss of accuracy. Those are evaluations the companies described to OpenAI. The post does not include the prompts, the sample sizes beyond Box’s 30 cases, or an outside audit.
Output tokens are off the bill. Cache writes are not
The price on the product post is $0.10 per million input tokens, and $0.01 per million for cached input. The post says callers pay only for input tokens, with no output-token charge. The guide repeats that: uncached input is $0.10 per million, cache reads take the discounted input price, cache writes are charged, and output tokens are not. On the standard price sheet, GPT-6 Luna’s short-context output is $0.50 per million tokens. That line is for other Luna requests. The Decisions guide says its rates apply to /v1/decisions, and that other requests using the same model follow the model’s own pricing. A team that sends the same text through the Responses API is on a different bill.
The cheap read has a floor. The guide says Decisions puts a cache breakpoint at the end of the shared input by default, and that the reusable prefix has to be at least 1,024 tokens. Text inside the questions, including images attached to a question, does not count toward that 1,024 and is outside the prefix that gets reused. Cache-write rates match Luna’s standard prompt-cache rates. On the price sheet, a short-context write kept for 30 minutes is $0.125 per million tokens, and a write kept for 12 hours is $0.25. The same sheet lists Luna’s long-context input at $0.20 per million, against $0.10 for short context. The Decisions guide says long-context multipliers apply. It does not, on that page, print the token count where short context ends.
Where the call runs is two different phrases. The product post says inference residency is available in the US and the EU, and that OpenAI has added capacity in several regions in Asia to improve latency. The guide is narrower on processing and wider on storage: regional processing is supported in the United States and in Europe, defined there as the EEA plus Switzerland, while data residency is available in all supported API regions. Zero Data Retention and HIPAA use are available to eligible customers. The Asia line in the blog is capacity. It is not listed as a regional-processing option. Voice is a separate hook. The post says Decisions can be called from a GPT-Live session, including on a transcript update, so an application can pick a next step while someone is still talking. That is the product claim. The latency numbers attached to it are Speak’s.


