OpenAI Launches Decisions API: 10x Faster Routing and Classification
OpenAI released the Decisions API in public beta. It evaluates text and images 10x faster than standard responses using gpt-6-luna.
Tools · Source: Hacker News
What happened
OpenAI just released the Decisions API in public beta. It evaluates text and images to return typed answers ten times faster than the standard Responses API. General availability is expected in the coming weeks. The only supported model right now is gpt-6-luna. You access it through a dedicated post request endpoint.
The architecture of a request is broken into three distinct parts. You define the model, provide the shared input evidence, and supply an array of questions to evaluate. The input can be a simple text string or a complex array of user messages containing both text and images. The API processes this shared context and returns an array of answers mapped to your unique question names.
OpenAI restricts the output to three strict question types. The predicate type checks a condition and returns a probability estimate from zero to one. The choice type selects a single option from a fixed set of supplied values. The score type rates an input against ordered levels and returns a probability-weighted average of the level indices. You can test all of these configurations in the OpenAI Playground before writing any actual code.
Key facts
- 10x — Speed increase over the Responses API
- gpt-6-luna — The only model currently supported for this API
- Python 3.26.0 — Minimum Python SDK version required to use the endpoint
Why it matters
This fundamentally changes how we build routing and moderation layers in AI applications. Until now, developers forced large language models to output complex JSON schemas just to handle simple classification tasks. That approach was slow, expensive, and prone to generation errors. Now you can use a dedicated API to classify content, route requests, and prioritize work at ten times the speed. You stop wasting valuable tokens on text generation when your application only needs a binary decision or a simple category label. This is a massive win for application latency.
The second-order effect is a permanent shift in AI system architecture. Fast, deterministic routing will become the default front door for all agentic workflows. You will put this Decisions API directly in front of your heavy reasoning models. It will act as a high-speed traffic cop for your infrastructure. It will instantly decide if a user prompt requires a heavy reasoning model, a specialized multi-agent workflow, or just a simple deterministic function call. This lowers costs and speeds up the user experience across the board.
For builders
Replace Structured Outputs for simple routing
Stop using the Responses API with Structured Outputs just to pick a category. Swap to the Decisions API using the choice question type for your routing logic. You will save massive amounts of latency and compute costs while getting more reliable outputs.
Build visual moderation instantly
The API accepts image inputs alongside text natively. You can use the predicate question type to check user-uploaded photos for visible damage or policy violations. It returns a clean probability score from zero to one, making visual moderation trivial to implement.
Upgrade your SDKs immediately
You cannot use this new endpoint with outdated libraries. You must upgrade your development environment to Python 3.26.0, Node 7.30.0, Go 3.73.0, Ruby 0.101.0, or Java 4.78.0. Update your package dependencies immediately before trying to implement this in production.
My take
I have been waiting for this exact primitive since we started building agentic workflows. Forcing generative models to do basic classification was always an expensive hack. OpenAI finally built a fast, dedicated traffic cop for our applications. This is going to make our products feel instantly snappier and drastically reduce our inference bills.
Original reporting: Hacker News. This is my rewrite and opinion.