OpenAI has made its Decisions API generally available, giving developers a dedicated way to make fast, structured decisions inside applications and AI-agent workflows. Powered by GPT-6 Luna, the API is designed to classify information, select between predefined actions and evaluate inputs against scoring criteria. OpenAI says it can return these decision results up to 10 times faster than using the same model through the Responses API.
The launch targets a specific challenge in modern software: applications often need to make hundreds or thousands of small decisions before completing a larger task. These include routing customer requests, categorizing transactions, checking content and determining the next action an AI agent should take. By providing focused outputs instead of requiring a general-purpose written response, the Decisions API aims to reduce latency and costs while helping developers build more responsive applications.
Key Takeaways
- General availability: OpenAI made the Decisions API generally available on October 22, 2026.
- Up to 10x faster: The API is designed for fast, bounded decisions compared with GPT-6 Luna through the Responses API.
- Three output types: Developers can request a probability, a choice from predefined options or a score against a rubric.
- Multimodal support: Inputs can include text, images and structured JSON data.
- Lower input costs: The published Decisions API price starts at $0.10 per million uncached input tokens, with cached input at $0.01 per million tokens.
- New capabilities: The general-availability release adds features including prompt caching, native GPT-Live integration and support for up to 400 questions per request.
What Is OpenAI’s Decisions API?
The Decisions API is a developer interface that turns information into structured answers an application can use to determine what happens next.
A conventional AI interaction might ask a model to read a customer complaint, explain its meaning and recommend a department. That approach can be useful when a detailed explanation is required, but it may be unnecessarily expensive or slow when the application only needs the department name.
The Decisions API lets the developer define the task and specify the acceptable outputs. Instead of producing an open-ended response, GPT-6 Luna evaluates the supplied evidence and returns an answer in the requested format.
For example, an e-commerce company could ask whether a product image shows visible damage. A customer-support platform could select the appropriate team for a complaint. A coding agent could score a proposed command’s risk before deciding whether it should be executed.
The API is available through the dedicated /v1/decisions endpoint and currently supports GPT-6 Luna as its model. Developers can experiment with decision questions in the Decisions Playground before integrating them into an application.
How the Decisions API Works
Developers provide three main elements: the information to evaluate, the questions to answer and the allowed output format.
The API supports three decision types.
| Decision type | What it returns | Example use case |
|---|---|---|
| Predicate | A probability that a condition is true | Estimate whether a package appears damaged |
| Choice | One selection from predefined options | Route a support ticket to billing or technical support |
| Score | A score based on ordered levels | Rate an issue as minor, moderate or severe |
A predicate is useful when an application needs an estimate of whether something is true. A choice is suitable for selecting a category or action from a fixed set. A score is designed for ordered evaluations, such as assessing severity against a defined rubric.
For choice and score questions, the API also provides probability information and a confidence measure. Developers can use these results to decide whether to proceed automatically or send uncertain cases to a human reviewer.
That distinction is important. A structured response makes the output easier for software to process, but it does not guarantee that the underlying decision is correct. Developers still need to test their questions, evaluate accuracy and choose appropriate thresholds.
Why OpenAI Says It Can Be 10 Times Faster
The central benefit is that the Decisions API focuses model processing on questions with bounded answers.
When a developer uses a general-purpose model interface for a simple classification task, the application may request a response that contains more information than it needs. A dedicated decision interface is designed to return the required result directly.
OpenAI says the Decisions API can deliver results up to 10 times faster than GPT-6 Luna through the Responses API. This is a maximum stated comparison, not a guarantee that every request will be 10 times faster. Actual latency depends on factors such as input size, question complexity, network conditions and workload.
The speed advantage can become particularly valuable in applications that make frequent decisions. If an AI assistant must classify a message, select a tool, evaluate a result and determine its next action, reducing the delay at each step can make the entire experience feel more responsive.
The API is therefore aimed less at replacing every interaction with a language model and more at accelerating the repetitive decisions that occur inside larger workflows.
New Features in the General-Availability Release
OpenAI introduced the Decisions API in public beta on October 6, 2026. Its October 22 general-availability announcement included several additions intended to make the service more useful in production applications.
Prompt caching: Applications can reuse shared input across requests and receive discounted pricing on cached input tokens. This is useful when many decisions rely on the same instructions or reference material.
Structured JSON input: Developers can pass structured data directly rather than converting everything into plain text first. This can simplify workflows involving transaction records, support tickets and business data.
Image support: Applications can evaluate images alongside text. Developers can also include images in individual question instructions, allowing separate questions to examine different visual evidence.
More questions per request: The API now supports up to 400 questions in a single request, double the previous public-beta limit of 200. This can help when multiple independent checks need to be performed on the same input.
Native GPT-Live integration: Live voice applications can trigger decisions as transcripts arrive, enabling an application to select a response or action while a conversation is in progress.
The release also includes additional capabilities such as conversation-history evaluation, usage reporting and private network connectivity for eligible configurations.
How Businesses Could Use the Decisions API
OpenAI’s announcement describes adoption across internal workflows and customer applications. The examples demonstrate how structured decisions can be integrated into systems that already use AI.
| Industry or workflow | Potential application |
|---|---|
| Customer service | Classify requests, route tickets and determine whether escalation is needed |
| Personal finance | Categorize transaction descriptions and organize spending |
| Enterprise software | Classify documents and direct them to the correct workflow |
| Cybersecurity | Score the risk of a proposed command before execution |
| Voice applications | Select an appropriate response or action during a live conversation |
| Education | Make language-learning interactions respond more quickly to user input |
OpenAI cited several early customer examples. Box reported roughly five-times-faster route selection in an evaluation involving 30 certificates of insurance, with accuracy matching its previous Responses API approach. Datadog reported 40-times-faster 95th-percentile latency for a command-risk scoring workflow compared with available frontier models, with comparable accuracy.
These are customer-reported results from specific evaluations, not universal benchmarks. Performance and savings will vary by application and implementation.
Decisions API Pricing and Availability
OpenAI has published the following base pricing for the Decisions API using GPT-6 Luna.
| Pricing component | Price |
|---|---|
| Uncached input | $0.10 per 1 million tokens |
| Cached input | $0.01 per 1 million tokens |
| Output tokens | No separate output-token charge for Decisions API |
Cache writes, regional processing premiums and long-context pricing rules can affect the final cost. Developers should consult the official pricing and availability guide before estimating production expenses.
The service is generally available through OpenAI’s API. Eligible customers can also use supported data-protection features, including Zero Data Retention, subject to the applicable requirements and limitations.
For developers already using the Responses API, the main decision is whether a task needs a generated explanation or simply a reliable structured result. The Responses API remains appropriate for open-ended generation, tool use and more flexible outputs. The Decisions API is designed for narrower classification, selection and scoring tasks.
The Bigger Picture
The Decisions API reflects a wider shift in AI software development: applications increasingly need models to make small operational decisions, not just produce text. These decisions may determine which team receives a ticket, whether a transaction needs review or which tool an AI agent should call next.
A dedicated interface for these tasks could help developers reduce response times and control costs in high-volume workflows. It also gives engineers a clearer way to evaluate the behavior of individual decision points rather than treating an entire AI interaction as one open-ended response.
The opportunity comes with a responsibility to test results and establish safeguards. A fast but incorrect decision can still cause financial, operational or security problems. Developers will need to measure accuracy, manage uncertainty and ensure that consequential actions remain subject to appropriate controls.
FAQs
1. What is OpenAI’s Decisions API?
The Decisions API is a developer interface that uses GPT-6 Luna to evaluate text, images and structured information, returning probabilities, predefined choices or scores that applications can use to make decisions.
2. Is the Decisions API 10 times faster than GPT-6 Luna?
OpenAI says the API can deliver decision results up to 10 times faster than using GPT-6 Luna through the Responses API. This is a stated maximum, and actual results depend on the workload.
3. How much does the Decisions API cost?
Published base pricing is $0.10 per million uncached input tokens and $0.01 per million cached input tokens. There is no separate output-token charge for Decisions API responses, although other pricing conditions may apply.
4. What kinds of data can the API process?
The API supports text, images and structured JSON input. Developers can use it to classify information, select among predefined options and score inputs against ordered criteria.
5. Should developers replace the Responses API with the Decisions API?
Not for every task. The Decisions API is suited to focused classification, selection and scoring. The Responses API remains useful when an application needs open-ended text generation, broader reasoning workflows or flexible tool interactions.
Looking Ahead
The next phase will depend on how developers apply the Decisions API in production and whether its speed and pricing advantages hold across different workloads. Applications with frequent, repeatable classification tasks may be among the clearest candidates, particularly when response time directly affects the user experience.
For OpenAI, the launch adds another specialized building block to its API platform. As AI agents become more integrated into software products, interfaces designed for particular stages of a workflow could help developers build systems that are faster, more predictable and easier to evaluate.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
