OpenAI Decisions API Is Live: Agents That Check the Rules Before They Act

Most AI products answer in paragraphs. OpenAI's Decisions API answers in numbers, choices and scores, and it does it fast. On 22 October 2026, OpenAI made the Decisions API generally available, pricing it at $0.10 per million input tokens with no charge for output tokens, and positioning it as the tool for the thousands of small, bounded decisions that slow down real applications.

Those small decisions are everywhere in production software. Is this support ticket urgent? Which department should handle it? Does this photo show visible damage? Is this transaction suspicious? Traditionally, answering each one with a full language model meant paying for a paragraph of reasoning and waiting a second or more for it. The Decisions API is built for the opposite: a single, structured answer, in about 150 milliseconds, at a price that makes high-volume use affordable.

The API first appeared in limited preview at OpenAI's DevDay on 29 September 2026, followed by a public beta on 6 October. General availability, a little over three weeks later, is therefore not a brand-new product. It is the moment OpenAI is telling developers the service is ready for production workflows, with new features and a stable price.


Credit: OpenAI logo, via Logodix

What the API actually does

The Decisions API takes three things: a model, an input, and a set of questions. The model is currently only GPT-6 Luna. The input can be text, images or native JSON, or a conversation history. The questions are the interesting part, because each one has a defined answer set.

There are three question types. A predicate returns a probability from 0 to 1 that a stated condition is true, for example whether a photo shows visible damage. A choice picks one value from a fixed list and returns a confidence plus a probability for every option, for example which department handles a ticket. A score returns a probability-weighted average of rubric level indices, for example grading a piece of work against a rubric.

The response is an answers array keyed by the question names. There is no paragraph to parse, no free-form text to interpret, and no need to write code that extracts a decision from a model's prose. That structure is the point. It turns a language model into a component that fits into ordinary software, the way a database query or a rules engine does.

Why speed and price matter together

OpenAI says the endpoint runs up to 10 times faster than routing the same bounded decisions through the standard Responses API with GPT-6 Luna. In demonstrations, latency came in at approximately 150 milliseconds. Nikunj Handa, OpenAI's API product lead, was reported as saying the service shrank response times from one to three seconds down to 100 to 180 milliseconds.

That speed comes from constraining the output. A general-purpose API call has full token generation overhead, because it must produce a whole response. Decisions constrains the output space to a probability, a choice or a score, and optimises accordingly. The result is not a smarter model. It is the same model, used in a narrower, faster way.

Pricing follows the same logic. At $0.10 per million input tokens, with cached input at $0.01 per million, and no output-token charges, the cost of a high-volume classification workload becomes predictable. For 1 million classifications of 500 tokens each, one analysis put the Decisions API cost at $50, compared with $21 for a competing decision model, which shows the API is not the cheapest option on every workload. But for teams already using OpenAI, the combination of speed, structure and predictable pricing is the selling point.

What changed between beta and general availability

General availability added several features that matter for production use. Native GPT-Live integration means a voice session can call Decisions as the user speaks, so a voice app can react in real time rather than waiting for a full turn to finish. Prompt caching helps lower costs when context is reused, though one guide warned that cache writes can cost more per token than a plain uncached call, because GA introduced prompt caching as a new feature the beta never had.

Native JSON inputs mean structured app data can be passed directly, without converting it to prose first. Images can now appear in individual question instructions, so multiple images can be compared against a shared reference. And the question cap doubled, from 200 in the beta to 400 per request, so one call can now carry twice as many decisions.

On compliance, the API supports Zero Data Retention for eligible customers and is HIPAA eligible, with inference residency available in the US and the European Union. Those details matter for healthcare, finance and other regulated uses, where a fast API is useless if it cannot meet data rules.


Credit: API interface illustration, via Quickwork

Who it is for, and who should skip it

The API is for teams with high-volume, repeated decisions: routing support tickets, classifying content, scoring outputs, prioritising work, moderating uploads, or checking documents against a rubric. Early trial customers reported by OpenAI include Box, Datadog, Ramp and Speak, which fits that pattern: companies with large volumes of structured work.

It is not for open-ended work. If you need a draft, an explanation, a summary or a creative response, this is the wrong tool. It is also not automatically cheaper than a purpose-built classifier. A specialised decision model may still win on price for a single, stable task. The Decisions API wins when the task changes often, when the input is messy text or images, or when you need a language model's judgement without its prose.

One pricing wrinkle deserves attention before production use. The headline rate is $0.10 per million uncached input tokens, and cache reads are cheap, but cache writes have their own pricing that does not show up if you only skim the headline rate. Teams should model their real mix of cached and uncached traffic before estimating costs, rather than multiplying the headline rate by total tokens.

What this says about where AI APIs are going

The Decisions API is part of a wider shift: from general-purpose chat models to specialised interfaces that do one job well. OpenAI already sells Luna as a fast, cost-efficient model for everyday work. Decisions takes that model and constrains it further, trading generality for speed and structure.

That shift matters for developers because it changes how AI is priced and used. Instead of paying for paragraphs and parsing them, teams can pay for decisions and use them directly. Instead of waiting seconds for a model to think out loud, they can get an answer in milliseconds. The model is the same. The interface is what changed, and the interface is often what decides whether AI is usable in production at all.

Credit: Web API development photo, via CodingFlow

Key Takeaways

  • OpenAI made the Decisions API generally available on 22 October 2026, after a DevDay preview on 29 September and a public beta on 6 October.
  • The API runs on GPT-6 Luna only, via POST /v1/decisions, and returns typed answers: predicates (probability 0 to 1), choices (fixed list with confidence), and scores (rubric-weighted average).
  • Pricing is $0.10 per million uncached input tokens and $0.01 per million cached input tokens, with no output-token charges.
  • One request can carry up to 400 questions, doubled from the beta, and inputs can be text, images, JSON or conversation history.
  • OpenAI claims up to 10 times faster performance than GPT-6 Luna via the Responses API, with demonstrated latency around 150 milliseconds.
  • New at GA: native GPT-Live integration, prompt caching, native JSON inputs, images in question instructions, Zero Data Retention for eligible customers, and US/EU inference residency.

The Decisions API will not write your next report or brainstorm your next product. It will answer, in milliseconds and in a format your code can use, whether something is true, which option fits, or how well something scores. For the thousands of small decisions that make up real software, that is exactly the right job, and OpenAI has now priced it for production.

Sources

Comments

Post a Comment