OpenAI Makes the Decisions API Generally Available for Fast, Bounded AI Choices
OpenAI [announced general availability on October 22](https://openai.com/index/decisions-api/) for the **Decisions API**, a dedicated endpoint for narrow questions whose answers fit a predefined form. The product was previously described in limited preview during DevDay; the news now is its wider release, updated capabilities and published pricing—not the invention of classification by AI.
What does the API do?
An app sends shared evidence as text, images or structured JSON to **POST /v1/decisions**, then supplies one or more questions. Each question asks for one of three answer types: a **predicate** returns an estimated probability, a **choice** selects from developer-defined options, and a **score** evaluates ordered rubric levels. The endpoint currently supports **GPT-6 Luna only** and up to **400 questions in one request**. OpenAI has also added prompt caching, question-specific images and a native GPT-Live integration since beta. [Sources: launch announcement](https://openai.com/index/decisions-api/); [developer guide](https://developers.openai.com/api/docs/guides/decisions)
For example, a support system could classify a ticket as billing, technical or shipping before routing it. A coding agent could use a choice question to select among *continue*, *retry* and *ask for review* after inspecting a tool result. Those are possible workflows, not a claim that the API autonomously performs the selected action: **your application still executes or approves what happens next**. [Source: developer guide](https://developers.openai.com/api/docs/guides/decisions)
Price and performance: what is actually claimed?
OpenAI lists **$0.10 per million uncached input tokens** and **$0.01 per million cached input tokens**, with **no output-token charge**. Cache writes are separately charged, and regional or long-context pricing adjustments can apply. These prices are specific to the Decisions endpoint; a regular GPT-6 Luna request through another endpoint has its own rates. [Sources: announcement](https://openai.com/index/decisions-api/); [pricing and availability guide](https://developers.openai.com/api/docs/guides/decisions)
OpenAI says the bounded-answer endpoint can be **up to 10× faster** than GPT-6 Luna through the Responses API. That is OpenAI’s comparison, not an independent benchmark. Before putting it in a high-volume or safety-sensitive path, test representative examples for false positives, false negatives, latency and total billed tokens. The guide recommends setting thresholds from labeled data and the cost of each kind of mistake. [Sources: announcement](https://openai.com/index/decisions-api/); [developer guide](https://developers.openai.com/api/docs/guides/decisions)
Where does it fit—and where does it not?
Use Decisions for repeatable classification, routing and prioritization where the allowed outcomes are known in advance. If a task needs a written explanation or an object shaped to your own JSON schema, OpenAI points developers to **Structured Outputs with the Responses API**; for model-selected tool calls with arguments, use **function calling**. The Decisions API is a component in a workflow, not a replacement for a general-purpose response or an agent runtime. [Source: developer guide](https://developers.openai.com/api/docs/guides/decisions)
The [API reference](https://developers.openai.com/api/reference/resources/decisions/methods/create) documents the request and answer structures. OpenAI says regional processing is supported in the US and Europe, while Zero Data Retention is available to eligible customers; teams handling regulated data should verify their own eligibility and agreements before relying on those controls. **Bottom line:** The launch gives developers a lower-cost, bounded-answer option for frequent small decisions, but the useful comparison is against their own labeled cases and current routing logic—not a headline speed multiplier.