Analysis
OpenAI's Decisions API: input-only pricing for the small decisions inside agents
1AYM is an OpenAI Select Partner.
Published . Vendor facts checked on against each vendor’s own pages, listed at the end. Plans, prices and availability change. Neither OpenAI nor Anthropic has reviewed this analysis.
The short version
OpenAI's Decisions API became generally available on 22 October 2026. It costs $0.10 per million input tokens, with no charge for output. Data can be processed in the US or Europe, and eligible customers can get Zero Data Retention.
- What the Decisions API does
- Generally available on 22 October 2026, built on GPT-6 Luna, with three answer types and up to 400 questions a request. OpenAI says it is up to 10 times faster than GPT-6 Luna through the Responses API.
- Price and data terms
- $0.10 per million input tokens, $0.01 per million cached, and no charge for output. Regional processing in the US and Europe, and Zero Data Retention and HIPAA use for eligible customers.
The Decisions API: typed answers for the small calls inside a workflow
What was announced
On 22 October 2026, OpenAI made its Decisions API generally available. It runs on GPT-6 Luna and handles questions with a bounded answer. Typical ones are how to classify an input, where to route a request and which action to take next. OpenAI says it returns these answers up to 10 times faster than GPT-6 Luna through the Responses API.
Developers write a question and its allowed answers, then pass the context as text, images or JSON.
A predicate answer gives the probability, from 0 to 1, that a condition is true, such as visible damage on a product. With a choice answer, the model picks one of your options, such as continue, retry or ask for clarification. It adds probabilities and a confidence figure. Score answers rate the input against ordered levels, such as how severe the damage is.
Since the public beta, OpenAI has added prompt caching for context that repeats, US and EU inference residency, JSON inputs and images in question instructions. A request holds up to 400 questions, twice the beta limit. OpenAI has also added capacity in several regions in Asia.
The API works with GPT-Live, OpenAI's real-time voice service. In OpenAI's example, a support agent decides during a live conversation whether to ask a clarifying question, route the request or escalate it.
Inside OpenAI, the dots team uses it to cut response times. The Personal Finance team sorts transaction descriptions with it, and the Communications team classifies coverage by theme. OpenAI names Box, Datadog, Ramp, Sierra and Speak as customers.
OpenAI quotes Box, which uses Box Extract with the Decisions API to classify business documents and route them. In Box's evaluation, as OpenAI reports it, route selection was about 5 times faster than through the Responses API, with the same accuracy. Cases with lower confidence went to reviewers. The evaluation covered 30 certificate-of-insurance cases.
What it means for enterprises
Many steps in agents and back-office automation already have this shape. A claims step checks whether a claim is complete. A document step picks a queue, and a support step decides whether a chat needs a person.
With a typed answer, the application gets a fixed set of outcomes, each with a probability. Each decision is then easier to log, test and audit.
The probability on each answer lets a team decide which cases a person sees. A claims or document team can send anything under a set confidence to a person, as Box describes, and let the rest through. OpenAI's guide recommends setting that line from labelled examples, weighted by what a wrong yes and a wrong no each cost.
OpenAI's updated Usage Policies take effect on 20 November 2026, as our post on those rules sets out. They bar high-stakes actions in sensitive areas, such as credit, insurance and jobs, without a person reviewing them. On our reading, a decision call that declines a loan or an insurance claim likely needs a person behind it.
Before you switch it on
- Busiest steps
- List the classification, routing and escalation steps in your agents and automations, and how often each runs. The steps that run most often carry the most tokens.
- Labelled cases
- Collect past cases with known right answers for each step you plan to move. Use them to set the confidence threshold, as OpenAI's guide recommends.
- Human review
- Mark each decision that falls in an area such as credit, insurance or jobs. Keep a person on those, whatever the confidence figure says.
- Your own test
- Run the same cases through your current setup and the Decisions API, and compare speed and accuracy. The only published accuracy figure is Box's evaluation of 30 cases.
Price and data terms: input tokens only, US and EU processing
What was announced
OpenAI charges $0.10 per million input tokens for the Decisions API and $0.01 per million for cached input. Output tokens are not charged. The developer guide adds that cache writes are charged at GPT-6 Luna's standard caching rates, and that regional processing premiums and long-context multipliers apply.
OpenAI's pricing page lists GPT-6 Luna on the standard tier at $0.10 per million input tokens and $0.50 per million output tokens. Cached input costs $0.01 per million. A 30-minute cache write costs $0.125 per million tokens, and a 12-hour cache write $0.25.
Regional processing is offered in the United States and in Europe, which the guide defines as the EEA plus Switzerland. Data residency is offered in all supported API regions. Zero Data Retention and HIPAA use are available to eligible customers.
The pricing page adds a 10% uplift on regional processing endpoints. It applies to models released on or after 5 March 2026 that are eligible for data residency. None of the three pages says whether that uplift applies to the Decisions endpoint.
What it means for enterprises
On OpenAI's published prices, input costs the same $0.10 per million tokens as GPT-6 Luna through the Responses API. The saving is on output, $0.50 per million tokens on the standard tier, plus whatever the speed gain is worth to you.
Our own illustration: a step runs a million times a month, with 2,000 input tokens and a 50-token answer through the Responses API. That is $200 of input and $25 of output a month at standard prices. The same input through the Decisions API costs $200, before caching or any regional premium, so the price saving is $25 a month.
A step with short answers saves little on price, so the case for moving it likely rests on speed and the typed answer. A step that writes longer output saves more.
Only US and European processing are named. A buyer whose policy requires UK or Gulf processing should ask OpenAI before planning production.
Before you switch it on
- Token mix
- Measure input and output tokens for each step you plan to move. The output share tells you how much of the bill the Decisions API removes.
- Caching
- Put shared instructions first, so they form a cacheable prefix of at least 1,024 tokens. Cached input costs $0.01 per million tokens against $0.10.
- Regional price
- If you need US or European processing, ask OpenAI whether the 10% regional uplift applies to the Decisions endpoint before you budget.
- Retention and health data
- Confirm your account is eligible for Zero Data Retention, and for HIPAA use where it applies, before you send personal or health data.
Where the market is heading: narrower, cheaper calls for the routine steps in an agent
OpenAI presents the launch as a fix for the cost and delay of the many small decisions an app or agent makes in a workflow. Its answer is a narrow endpoint on an existing model, GPT-6 Luna, priced on input only.
The guide says the API returns typed answers only, and your application carries out any action that follows. For objects in your own JSON schema or for tool-call arguments, OpenAI points to Structured Outputs and function calling.
For a buyer, that likely means agents built from more than one kind of call. A general model drafts and plans, and a typed decision call handles the routine yes, no and route steps. Each step can then be priced, tested and reviewed on its own.
| Dimension | Where it stands | Said by |
|---|---|---|
| Status | Generally available since 22 October 2026 | OpenAI announcement |
| Model | GPT-6 Luna only | Developer guide |
| Answer types | Predicate (probability from 0 to 1), choice (one of your options), score (2 to 10 ordered levels) | Announcement and guide |
| Input price | $0.10 per million tokens; cached input $0.01 | Announcement and guide |
| Output price | Not charged (GPT-6 Luna through the Responses API: $0.50 per million) | Guide and pricing page |
| Questions per request | Up to 400 (200 in the beta) | Announcement and guide |
| Inputs | Text, images, JSON and conversation history; no files or audio | Developer guide |
| Regional processing | US and Europe (EEA and Switzerland) | Developer guide |
| Regional price | Premiums apply; the rate for this endpoint is not stated | Guide and pricing page |
| Zero Data Retention and HIPAA use | Eligible customers | Developer guide |
| Rate limits | Not stated | None of the three pages |
What is not known yet
We read the four sources on 23 October 2026, the day after the launch. None of them answers the points below.
- Rate limits
- OpenAI hasn't published rate limits for the Decisions endpoint on the pages we read.
- Regional price
- The guide says regional processing premiums apply. It doesn't say whether the pricing page's 10% uplift covers the Decisions endpoint.
- UK and Gulf processing
- Only US and European processing are named. Nothing says whether UK or Gulf processing is planned.
- Accuracy
- The only accuracy figure is Box's evaluation of 30 cases, as OpenAI reports it, and no independent figures are published yet.
- Speed
- OpenAI doesn't say how it measured the claim of up to 10 times faster, or on which workload.
For engineers: the endpoint, answer types, inputs, caching and thresholds
Requests and answer types
Calls go to POST /v1/decisions, and gpt-6-luna is the only supported model. A request holds up to 400 questions, up from 200 in the beta. You can try questions in the Decisions Playground before writing code.
A predicate returns a probability from 0 to 1. For a choice question, the response is one of your predefined options with probabilities and a confidence value. Score questions need 2 to 10 ordered levels, and the result is the probability-weighted average of the level indices, so it can fall between levels.
Inputs and caching
Context can be text, images as base64 or public URLs, and JSON through input_json in user messages. Conversation history, including tool outputs, is accepted. file_id inputs, files, audio and item references are not.
A cacheable shared prefix must be at least 1,024 tokens. Put the instructions and fixed context first and the case-specific input last.
Sources: [2]
What it doesn't do
It returns typed answers only. For objects in your own JSON schema or tool-call arguments, use Structured Outputs and function calling. Questions that depend on earlier answers need separate requests, so a chain of decisions means several round trips.
Sources: [2]
Setting thresholds
The guide treats the probabilities as estimates. Set each threshold from labelled examples, weighted by the cost of a false positive and of a false negative.
Log the probability and the threshold with every decision. A reviewer can then see why a case went through or went to a person.
Sources: [2]
Where 1AYM fits
If you are moving high-volume classification, routing or escalation steps onto OpenAI's Decisions API, bring in 1AYM. We are an OpenAI Select Partner and model-selective across OpenAI, Anthropic, Gemini and open-source. We pick the model for each workload from production evidence.
Our Enterprise AI Platform & Agentic Workflows engagement builds workflow agents and the middleware underneath them. It builds in deterministic verifier gates on consequential actions and evaluation running in CI. The gates produce their own audit trail, and human approval sits on the decisions that warrant it.
To choose which steps to move first and what the move saves, start with our AI Opportunity & Feasibility Sprint. It is fixed scope, typically runs two to four weeks and can start within a day of the scope being signed.
Sources
Frequently asked questions
Who can help us move routing and classification steps to the Decisions API?
1AYM, an OpenAI Select Partner. We build deterministic checks, thresholds and human approval sized to the risk, tested on representative data before anyone depends on it. Fixed-scope work can start within a day of the scope being signed. Book a 30-minute call or email enquiries@1aym.com.
How much does the Decisions API cost?
$0.10 per million input tokens and $0.01 per million cached input tokens. Output tokens are not charged. Cache writes are charged at GPT-6 Luna's standard caching rates, and regional processing premiums and long-context multipliers apply.
Is it cheaper than GPT-6 Luna through the Responses API?
Input costs the same, $0.10 per million tokens. The difference is output, which costs $0.50 per million tokens on the standard tier through the Responses API and nothing through the Decisions API. For calls with short answers the price saving per call is small, and speed is the larger gain if OpenAI's figure holds on your workload.
Where is our data processed?
Regional processing is offered in the United States and in Europe, which the guide defines as the EEA plus Switzerland. Zero Data Retention and HIPAA use are available to eligible customers. No UK or Gulf processing region is named.
Can it decide credit or insurance cases on its own?
From 20 November 2026, OpenAI's updated Usage Policies bar high-stakes actions in sensitive areas, such as credit, insurance and jobs, without a person reviewing them. The Decisions API can sort and score those cases for the person who reviews them.
Further
- OpenAI on Google Cloud and the 20 November rules · OpenAI's updated Usage Policies, including the human review rule for high-stakes decisions in areas such as credit and insurance.
- AI agent cost · A worked example of what an invoice-matching agent costs to build and run, and why upkeep costs more than the model.
- Agent proposes, verifier gates · How rules, tests and a person decide whether a model's output goes ahead in finance and compliance work.
- AI workflow automation: what to automate first · Score each process on value, feasibility and risk before you choose the first one to automate.
- Claude Haiku 5.5 pricing for enterprises · Anthropic's new small-model price, and what decides how much of it an enterprise keeps.
Running thousands of small decisions inside your agents?
Half an hour is enough to pick the steps worth moving, check what the move saves and agree where a person still reviews.
Last reviewed · 1AYM