OpenAI Decisions API Turns Evaluations Into Quick Verdicts

OpenAI has added a new tool to its developer platform that does one job: it makes a call. The Decisions API takes text, images, or a mix of both and returns a narrow verdict instead of a free-form answer. The company paired the launch with a simpler structure for its paid API tiers.

The Decisions API is in public beta now. OpenAI says general availability will follow soon.

What the Decisions API does

Most developers reach OpenAI models through the Responses API, which can generate open-ended output of almost any kind. The Decisions API is far narrower. According to OpenAI, it runs roughly ten times faster than the Responses API.

It supports three kinds of answers:

1. Yes/no probabilities. The model estimates how likely a statement is to be true for a given input.

2. Picks from a fixed list. The developer defines the categories in advance, and the model chooses one.

3. Scale-based ratings. The model places the input somewhere on a defined range.

None of the three formats lets the model wander. The developer sets the shape of the answer, and the model fills in the blank.

Where OpenAI sees it being used

OpenAI points to a few practical examples. One is spotting damage in photos. Another is routing incoming customer inquiries to the right place automatically. A third is classifying documents.

These are high-volume, repetitive tasks where a long written response adds little value. A support system does not need an essay about an email. It needs to know which queue the email belongs in.

For organizations handling sensitive material, the API supports zero-data retention, meaning inputs are not kept after processing. It can also be used in a HIPAA-compliant way in the US and Europe. HIPAA is the US law that governs how health information must be protected.

Price and model support

For now, only one model works with the Decisions API: gpt-6-luna. Input costs $0.10 per million tokens. Output tokens are free.

That pricing structure fits the product. A decision call produces very little output (a probability, a label, or a rating), so most of the cost sits on the input side anyway. Billing only for input makes costs easier to predict for teams running large numbers of classification jobs.

A response to "decision models"

The launch appears to be OpenAI's answer to a small but growing trend. Jev started a wave of so-called "decision models" in mid-September, and other players have followed with their own takes, including Cloudflare's fast decision models for agents and an open decision model from AWS. The idea behind all of them is similar: many steps in an AI workflow do not need a full generative answer, only a quick and reliable judgment.

Three tiers instead of five

Alongside the new API, OpenAI has trimmed its paid API tiers from five to three. They are now called Build, Launch, and Grow.

Organizations move up automatically once their total credit purchases cross the next threshold. Monthly usage limits are set at $500 for Build, $5,000 for Launch, and $200,000 for Grow.

Our Take

The Decisions API suggests that the big model providers are starting to treat narrow, structured judgments as a product category of their own, not just a prompt pattern developers build on top of chat models. That matters for anyone building agents or automated pipelines. A fast, constrained yes/no or category call is easier to test, cheaper to run at scale, and harder to push off course than an open-ended response.

It also fits a wider push on price, seen in moves such as Anthropic's steep Haiku price cuts. Free output tokens are an aggressive signal in that race.

It is worth watching whether OpenAI adds more models beyond gpt-6-luna before general availability, and how the speed claims hold up in real workloads. The simpler tier system, meanwhile, should make it easier for smaller teams to understand what they can spend before they hit a ceiling.