In this article
  1. What happened
  2. Why it matters
  3. What to do now
  4. What’s still unknown

TL;DR: On October 22, 2026, OpenAI made its Decisions API generally available. Instead of writing a free-text reply, it answers bounded questions about text, images or JSON with one of three typed results: a probability that something is true, one option from a list you supply, or a score on a rubric you define. It runs on GPT-6 Luna, and OpenAI says it is up to 10x faster than asking Luna the same thing through the Responses API. It costs $0.10 per million uncached input tokens and $0.01 per million cached input tokens, and output isn’t charged; cache writes are billed at GPT-6 Luna’s standard caching rates. If your app or automation already uses a model to sort tickets, tag transactions or decide “escalate or not,” this is worth a test this week.

What happened

OpenAI announced general availability on October 22, 2026, after a public beta. According to OpenAI’s announcement and developer guide:

  • What it does. You send shared evidence (a text string, a conversation, images, or JSON) plus a list of questions. Each question has a type. A predicate question returns a probability from 0 to 1 that a condition is true, such as “does this product photo show damage?” A choice question returns one of the values you supplied, such as billing, technical or shipping, with probabilities and a confidence figure. A score question rates the input against 2 to 10 ordered levels you define, such as issue severity.
  • The model and endpoint. Requests go to POST /v1/decisions, and gpt-6-luna is the only supported model.
  • What’s new since the beta. Prompt caching, images inside individual questions, native JSON input, up to 400 questions per request (up from 200), hosted image URLs, conversation history including tool calls, organization usage reporting, Private Link, and a native integration with GPT-Live, OpenAI’s model for voice applications. OpenAI also says it added capacity “in several regions in Asia” to cut latency.
  • Pricing. Uncached input costs $0.10 per million tokens and cached input $0.01 per million. Output tokens aren’t charged. The guide says cache writes are charged at the standard GPT-6 Luna caching rates, and that regional processing premiums and long-context multipliers still apply.
  • Data controls. Zero Data Retention and HIPAA use are available “for eligible customers,” and regional processing is available in the US and Europe (EEA plus Switzerland).

OpenAI published quotes from five early customers. Box says it saw about 5x faster routing of insurance documents with the same accuracy as the Responses API. Ramp says attributing its AI costs became 4x faster and 30% cheaper than with GPT-6 Luna. Sierra says it runs next-step choices in live customer conversations at under 280 ms p95. These are the companies’ own tests, not independent benchmarks.

Why it matters

Most small-business automations are full of small judgment calls: is this email a refund request, which person should get this lead, does this review need a reply today, is this receipt a business expense. Teams usually ask a chat model and then parse its text reply, which is slower, costs output tokens, and sometimes returns something your code didn’t expect.

Decisions changes three things for that kind of work:

  1. The answer is always one of your values. A choice answer is always one of the options you listed, so your Zapier, n8n or custom code can branch on it without parsing prose. Add an “other” option, as OpenAI suggests, for anything that doesn’t fit.
  2. You get a number to set a threshold on. Because answers come with probabilities, you can auto-handle the confident cases and send the borderline ones to a person. OpenAI’s guide says to pick thresholds using labeled examples from your own data.
  3. The cost structure favors classification. With no output charge and many questions per request, one call can tag a ticket for department, urgency and sentiment at once. If the same long instructions or reference document come first in every call, cache hits cut that part of the input to a tenth of the price, though writing the cache is charged at GPT-6 Luna’s standard caching rates. The prefix must be at least 1,024 tokens to be cached. We covered how OpenAI’s longer prompt caching pays off on October 21.

It is not a replacement for everything. OpenAI says to keep using Structured Outputs in the Responses API when you need to extract fields or get a written explanation, and function calling when the model should request a tool call. Decisions only returns the answer. Your app still takes the action.

What to do now

  1. Pick one high-volume decision. Good candidates are inbox triage, lead routing, review flagging and expense categories. Write it as a question with a short, clear description for each option.
  2. Build a small labeled test set. 30 to 50 real examples where you already know the right answer is enough to compare Decisions with what you use today on accuracy, speed and cost. OpenAI’s Playground lets you try questions before writing code.
  3. Set a review threshold. Send low-confidence answers to a person instead of acting on them automatically, especially for anything touching money or customers.
  4. Check your caching. Put the unchanging material (your rules, product list or reference image) first in input and keep it identical across calls so the cache can be reused. Question text is never cached.
  5. If you build voice agents, the GPT-Live integration lets a live call trigger decisions as the transcript arrives, for example to decide when to escalate. For background, see our earlier piece on OpenAI’s prompt-to-voice API.

What’s still unknown

OpenAI’s comparisons are against its own Responses API and come from OpenAI and selected customers. We haven’t seen independent accuracy tests. GPT-6 Luna is the only model for now, and OpenAI hasn’t said whether other models will be added. The exact cache-write prices are on OpenAI’s pricing pages rather than in the announcement.

Sources & method

Sources

Method. Last checked October 23, 2026, at 7:30 a.m. Manila time (23:30 UTC on October 22). Features, limits, prices and data controls come from OpenAI’s announcement and developer guide. Customer results are quoted from OpenAI’s announcement and haven’t been independently verified. The use cases and the steps under “What to do now” are our own suggestions; we haven’t run our own benchmark. No affiliate links on this page. How we work.

Spotted an error? See our corrections policy and tell us.