OpenAI made its Decisions API generally available on 22 October 2026. Instead of writing a reply, it answers a fixed question about text, images or JSON with a probability, a choice from your list or a score, so software can sort, route and flag work on its own. It runs on GPT-6 Luna, input costs $0.10 per 1M tokens and output is not charged. For a small business, the value is in routine jobs: tagging enquiries, checking photos, queueing tickets.
What happened
OpenAI's API changelog entry for 22 October 2026 says the Decisions API is now generally available with gpt-6-luna. The same changelog shows the beta started on 6 October 2026, when OpenAI described it as a way to turn text and images into typed answers 10x faster than its Responses API, the endpoint most developers use to get written replies from a model.
TechCrunch reported on 22 October 2026 that the Decisions API, first announced at OpenAI's developer day on 29 September, became available to all customers that Thursday.
The release adds several things the beta lacked. You can now send up to 400 questions in one request, up from 200 in the beta. Shared input can be cached for a discount, each question can carry its own image, and structured data can be passed directly as JSON. OpenAI's Decisions guide also lists conversation history, usage reporting and Private Link, a private network connection.
According to TechCrunch, OpenAI did not train a new model for this. Its API team took Luna and reconfigured it to run faster and cost less. Nikunj Handa, product lead for OpenAI's APIs, told the outlet that response times had fallen from one to three seconds to around 100, 120 or 180 milliseconds.
Key details
| Item | What OpenAI's pages say |
|---|---|
| Status | Generally available from 22 October 2026; beta from 6 October 2026 |
| Model | gpt-6-luna, the only supported model |
| Question types | Predicate (a probability from 0 to 1 that something is true), choice (one of your listed options) and score (a rating on 2-10 ordered levels) |
| Input | Text, images and JSON; files, audio and item references are not supported |
| Questions per request | Up to 400 (200 in the beta) |
| Price | Uncached input $0.10 per 1M tokens; output tokens not charged; cache writes charged; regional processing premiums and long-context multipliers apply |
| Data controls | Zero Data Retention and HIPAA use for eligible customers; regional processing in the United States and Europe (EEA and Switzerland) |
Why it matters
Most business uses of AI have been about writing. Much office work is deciding instead: is this email a complaint or a sales enquiry, is this delivery photo showing damage, is this ticket urgent. The Decisions API is built for that work.
The examples in OpenAI's guide make the point. One checks a product photo for visible damage such as a crack, tear or dent. Another routes a customer complaint to billing, technical, shipping or "other". A third scores how damaged a package is on three levels: none, minor or major. Each returns a number your software compares against a limit you set.
It is also a sign of a wider shift. We covered Microsoft-Decision-1 when Microsoft released it, and TechCrunch describes a whole category of these "decision" models now competing on speed, price and accuracy. In the two public rankings TechCrunch cites, the AIM System One Index and the Artificial Analysis Decision Index, OpenAI's model placed eighth and fourteenth. The article adds that results depend on the workload, so test on your own data.
| Decision model | Where it runs | Published price |
|---|---|---|
| OpenAI Decisions API (gpt-6-luna) | OpenAI's API | $0.10 per 1M input tokens; output not charged |
| Microsoft-Decision-1 | Microsoft Foundry and OpenRouter | $0.042 USD per million input tokens; output free |
| gpt-6-luna for written replies | OpenAI's other API endpoints | Standard price $0.10 per 1M input tokens and $0.50 per 1M output tokens |
Speed matters more than it sounds. TechCrunch reports that Speak, one of the early trial customers, cut the time between actions in its product from 1.4 seconds to 0.13 seconds.
What this means for businesses
You do not call this API yourself. Your developer, software provider or the systems you already use would. Your part is spotting the repetitive sorting and checking jobs it could take off your team, and setting rules around them.
Good candidates are tasks where the possible answers are known in advance. If you run an online store in Brisbane, that might be tagging incoming emails as returns, delivery problems or product questions, or flagging customer photos that show a damaged item. A trades business might sort job requests by urgency. These are the same jobs that come up for retail and e-commerce businesses when the store, stock and accounting systems need to work together.
Before you or your provider build anything, work through this checklist:
- Write the question around something you can see. OpenAI's guide says to write questions around observable criteria and to give each choice a distinct meaning. "Does the photo show a crack, tear or dent?" works; "Is this customer unhappy?" is vague.
- Always include an "other" option. The guide recommends a fallback such as "other" that sends unclear cases to a general review queue, so nothing is forced into the wrong box.
- Set your limits with real examples. The guide says to use labelled examples from your own application to choose thresholds, based on the cost of a false alarm against the cost of a miss.
- Keep a person on decisions about people. business.gov.au says AI tools need human oversight and that AI-generated content and decisions should be checked before you use them. Hiring is the clearest case: when you screen applicants, a recruitment and candidate verification platform such as Talent Tabloid gives a person verified evidence to review, not just a score.
- Remember the API only answers. OpenAI says Decisions returns typed answers and your application runs any actions. Rules, logs and permissions belong in your own software. Our guide to AI agent permissions, data access and cost covers how to keep that under control.
- Check the data settings. Zero Data Retention is offered to eligible customers only, and regional processing covers the United States and Europe. If enquiries hold personal information, ask your provider which settings apply and where data is processed.
If your software already uses OpenAI models, also check which ones. OpenAI is retiring several older models from its API, as we explained in GPT-4o and GPT-4.1 leaving OpenAI's API; moving to newer models is a good moment to ask which tasks could be decisions rather than written replies.
Often the right answer is not new software at all. Before recommending a build we check whether an existing product, or a connection between the systems you already have, would solve the problem for less. Our business systems and integrations work covers the plumbing and custom software covers the rest. For a second opinion on where automated sorting fits, ask us for a free first consultation and a written quote before any work starts.
Frequently asked questions
What is the OpenAI Decisions API?
It is an OpenAI service that answers a fixed question about text, images or JSON with a typed result: a probability that something is true, one choice from your list, or a score on ordered levels. It runs on gpt-6-luna and became generally available on 22 October 2026.
How is it different from asking ChatGPT?
ChatGPT and the Responses API write answers in words. The Decisions API returns a number or a label your software can act on straight away, and OpenAI says it does this about 10x faster than the Responses API. For extracted fields or written explanations, OpenAI points developers to Structured Outputs instead.
What does the Decisions API cost?
OpenAI's guide says uncached input costs $0.10 per 1M tokens and output tokens are not charged. Cache writes are charged, cache hits are discounted, and regional processing premiums and long-context multipliers can apply, so ask your developer to estimate usage on your real volumes.
Can it make decisions about customers or staff on its own?
It should not be left to. OpenAI says the API only returns answers and your application runs the actions, and business.gov.au says AI decisions should be checked by people. Keep human review on anything that affects a customer's account, a job applicant or an employee.
Does a small business need a developer to use it?
Yes, or a software provider that builds it into a product you already use. It is reached from other software, not a chat window, so someone has to connect it to your email, store, CRM or help desk.
Key takeaways
- OpenAI's Decisions API became generally available on 22 October 2026, running on gpt-6-luna.
- It returns a probability, a choice or a score instead of written text, and handles up to 400 questions per request.
- Uncached input costs $0.10 per 1M tokens and output is not charged.
- The strongest fits are repetitive sorting and checking jobs with known answers, with an "other" option and a person reviewing decisions about people.
- Test on your own examples before you trust it: public rankings vary and depend on the workload.