OpenAI Launches Decision API in Beta at $0.10/M Tokens

OpenAI is offering a specialized API for asking closed questions to a model, with standardized responses and billing at $0.10 per million input tokens. According to the company, this endpoint can make decisions up to ten times faster than using the model via the responses API. The current beta relies on gpt-6-luna and accepts both text and images. This service is aimed at routing, scoring, and guided selection, rather than free generation.

Pricing, Additional Costs, and Announced Usage Limits

At launch, the decision API is priced at $0.10 per million input tokens, with no cost for output tokens or for read/write operations in the cache. However, additional coefficients are applied for extended contexts, as well as specific fees based on the processing region. For one million requests, each averaging 1,000 billable input tokens, the cost would amount to $100 based on the standard rate, taking into account both the questions and their descriptions in the calculation. The service does not replace free generation uses: a task that requires a written explanation or the extraction of an object with arbitrary fields must go through a generation interface. Attempting to resolve, for example, an invoice extraction via twenty classification questions is not appropriate. Finally, the choice of a confidence threshold is not prescribed: an illustrative threshold is not a recommendation, and it is advised to start from manually annotated data to measure errors at various thresholds before setting one's own.

Where to Use It: Routing, Research, and Agent Actions

The API targets specific decisions: directing a message to the correct queue, selecting the relevant search index, or choosing the next action of an agent from an authorized list. Compared to the JEV model from TypesafeAI, a highlighted advantage is the ability to process images. Inputs can thus combine text and media, which broadens the practical decision-making cases without free text generation.

What the API Expects and What It Returns

Each call provides a model, shared inputs, and a list of questions. Three types of questions are supported: predicate returns a probability between 0 and 1, choice returns a selected value with probabilities and confidence level, and score returns an ordered level with probabilities and confidence. A predicate is not a boolean: it is up to the client to apply their threshold. Responses arrive in the order of the questions, and the API may refuse certain questions, hence the need to check the type of each response before reading its fields. Requests are made to /v1/decisions, and responses integrate directly into application logic. Choices may include values and descriptions to clarify the scope of categories.

Performance and Technical Scope of the Beta

According to OpenAI, the decision API can make decisions up to ten times faster than using the model via the responses API. The current beta operates with the gpt-6-luna model and exposes a dedicated decision endpoint. The service is offered in public beta following a presentation at DevDay.

Setup on the Python Side: SDK, Key, and Environment

Support for the decision API is available in the openai-python client starting from version 3.26.0, which can be installed via pip with openai==3.26.0. An API key is required: you need an OpenAI account with a payment method and credit, accessible via https://platform.openai.com/home, where a link on the left allows you to create a secret key. The key must then be placed in the OPENAI_API_KEY environment variable, as shown by an example PowerShell command. The provided examples are standalone Python programs; one sets a client timeout of 20.0 seconds and illustrates the use of a confidence threshold of 0.8 to switch to manual triage, with queues for returns, billing, delivery, and general. Another demonstrates two checks on a document (presence of a deadline and a postal address), and imports RateLimitError. A return within 30 days is mentioned, followed by a truncated email address.