OpenAI speeds up agent routing with Decisions API

Every developer building complex AI agents knows the routing tax. You need a fast check to decide which tool to call or whether an action is safe, but waiting for an LLM to generate tokens slows down the entire user experience.

OpenAI just launched the Decisions API in public beta, rolling it out to all developers to fix that exact problem.

Powered by GPT-6 Luna, the new endpoint is designed for split-second choices. OpenAI says it operates up to 10x faster than running GPT-6 Luna through the standard Responses API.

Three ways to make a choice

The API accepts text and image inputs, but ditching standard conversation in favor of structured outcomes. It offers three distinct output formats.

The first is Predicates, which estimates the probability that a specific statement is true.

The second is Choices, letting the model select from a predefined list of options alongside confidence scores.

The third is Scores, which evaluates the input against a numeric range.

By narrowing down what the model returns, it fits cleanly into workflows like request routing, dataset categorization, visual analysis, form navigation, and flagging risky tool calls before execution.

Zero output token charges

The pricing structure is just as aggressive as the latency improvements.

With GPT-6 Luna, inputs run at $0.10 per 1 million tokens. The kicker is that developers only pay for input tokens.

There are no cache-read charges, no cache-write charges, and zero output token fees whatsoever.

Why it matters

Routing is often the quiet killer of agent performance. If your orchestrator takes several hundred milliseconds just to pick a downstream tool, your app feels sluggish no matter how smart it is.

We think OpenAI nailed the economics here. By charging only $0.10 per million input tokens with no output fees, they have effectively eliminated the cost hesitation around adding deterministic classification layers across an app.

For the broader industry, this puts heavy pressure on specialized routing startups and lightweight self-hosted classifier models. Why manage your own small models when OpenAI gives you frontier-level routing at near real-time speeds for pennies?

The risk, as we see it, is platform lock-in. Decisions API is not a generic text completion endpoint; it is a proprietary decision primitive that binds your system architecture tightly to OpenAI.