Skip to content

All articles

OpenAI makes Decisions API generally available for business apps

OpenAI has moved its Decisions API out of beta. The tool allows business applications to classify data and route tasks without paying for output tokens.

On October 22, OpenAI moved its Decisions API (opens in a new tab) out of beta and made it generally available. This tool allows software to classify text and images, evaluate inputs and determine next steps without generating a full model response. For a small or mid-sized business, this alters the cost of building artificial intelligence into daily operations. Instead of paying a language model to write out a thought process every time a customer sends an email, a company can use a lightweight model to simply make a routing choice. This reduces both the delay and the cost of automated workflows.

Moving away from text generation

Most businesses interact with artificial intelligence through text generation. A user sends a prompt, and the model replies with a paragraph. When companies build internal tools, they often force this text generation process to act as a router. They ask the model to read a support ticket and reply with a single word like a category name. This method is slow and expensive because the model processes the request as a creative writing task, generating tokens one by one.

The Decisions API changes this structure. It uses a specific model named gpt-6-luna. Instead of writing text, the interface is designed to handle high-frequency workflow logic within business applications. It accepts native JSON inputs, meaning developers can pass structured application data directly to the model without converting it into a conversational prompt first.

The system can process up to 400 questions in a single request. OpenAI noted this capacity allows developers to make twice as many decisions in a single call. This allows a company to evaluate a single piece of customer feedback against dozens of criteria at once, checking for urgency, sentiment, product mentions and routing requirements simultaneously.

The exact costs

The pricing structure reflects the shift away from text generation. OpenAI charges $0.10 per million input tokens for the Decisions API. Because the model does not generate conversational text, there are no output token charges at all. A business only pays for the data it sends to the system.

The interface also includes automatic prompt caching. When an application sends the same large block of context repeatedly, the system remembers it. For these cached inputs, the price drops to $0.01 per million tokens.

This caching mechanism is particularly useful for applications that process growing chat histories or analyze multiple questions against a single large document. Developers on the OpenAI forum noted that the cache applies to the main input rather than the individual questions. This means a business can upload a large policy document once, pay the initial input cost, and then run multiple cheap classification tasks against it throughout the day.

What it means for your business

For a company of 50 to 500 people, the Decisions API makes high-volume automation practical. Customer support teams receive hundreds of messages a day. Previously, using a language model to read and route every single message incurred a high token cost and introduced a noticeable delay.

A business can build software that classifies inbound content and routes customer support tickets using pre-defined structured outputs. The system reads the email, decides it belongs to the returns department, and forwards it in near real-time. OpenAI stated the tool is "ready to add near real-time decisions" to production workflows.

This capability also enables applications to determine an autonomous agent's next step without incurring the full latency or token cost of complete generation loops. If a company uses an internal artificial intelligence assistant to manage inventory, that assistant needs to make dozens of small choices every minute. It must decide whether to check a database, email a vendor, or alert a human manager. The Decisions API allows the assistant to make these choices instantly and cheaply.

Finance leaders will appreciate the predictable pricing. By eliminating output charges, a company can forecast its automation costs based entirely on the volume of incoming data. A logistics firm processing thousands of delivery updates daily can calculate exactly what the routing logic will cost, without worrying about a model generating unusually long responses that inflate the bill.

What to watch

The update includes native GPT-Live integration. This feature allows voice applications to react as people speak. A business building a voice-based customer service agent can use the Decisions API to monitor the conversation and trigger actions in the background. If a caller mentions a specific product, the system can instantly retrieve the manual before the caller finishes their sentence.

Companies should watch how their engineering teams adapt to the caching rules. To get the lowest price of $0.01 per million tokens, developers must structure their requests so the large context remains static while the short questions change. If an application constantly modifies the base document, it will trigger the higher non-cached rate of $0.10 per million tokens.

The update also allows images in individual question instructions. This means a system can compare multiple images against a shared reference. A quality control application could look at a master image of a perfect product and instantly evaluate hundreds of photos from the assembly line to spot defects.

The Decisions API turns artificial intelligence from a conversational partner into a piece of silent infrastructure. By stripping away text generation, OpenAI has provided a faster and cheaper way to route data. Businesses that adopt this approach will process their daily workflows with much greater efficiency.