OpenAI’s Decisions API: The $0.10-per-Million Endpoint That Replaces Your Jury-Rigged Classifiers

What it is

OpenAI opened its Decisions API in public beta on October 6, 2026, and it is the most useful boring endpoint it has shipped in a while. Instead of asking a model to write prose, you ask it to decide: is this true (predicate), which of these options is right (choice), or how does this rate on a scale (score). It runs only on gpt-6-luna, costs $0.10 per million input tokens with no output, cache-read, or cache-write charges, and OpenAI says it answers about 10x faster than the Responses API. If your codebase is full of “is this spam / which team / should this go through” calls routed through a chat model and a regex, this is the endpoint that replaces them. [E1 our take]

The three call types

A predicate returns a probability between 0 and 1 that a condition holds. A choice picks one option from your fixed list and returns per-option probabilities. A score gives a probability-weighted rating on an ordered rubric [E3 OpenAI Decisions API docs]. You can ask multiple independent questions against one input in a single request, and there is an explicit refusal answer type. Inputs are text and images — but images must be inline base64 data URLs; hosted links and file IDs are refused [E4 AI Weekly / codeyourcraft, Oct 6 2026]. OpenAI’s own examples: inspect a product photo for damage, route a complaint to the right department, rate risk or severity.

Pricing that changes engineering habits

The hook is input-only billing at $0.10 per million tokens, nothing for output [E3 OpenAI pricing]. Most production systems are full of small decisions that sit in front of the expensive model, and teams usually handle them with a general chat model and a prompt begging for a one-word reply, then pay for output tokens and parse the text. Decisions turns that into a typed call where cost depends almost entirely on how much evidence you send [E4 community benchmarks

Bottom line

Decisions is a small, unglamorous release that fixes a real pattern: most production AI systems bury dozens of trivial classify, route, or flag calls behind a general model and a fragile prompt. If you are paying for output tokens to get a one-word answer, this endpoint is almost certainly cheaper and more reliable. Use it for high-volume, typed, low-creativity decisions; keep the big model for the work that actually needs judgment. [ E1 our take]

More in this category

• LLM tokens & pricing hub — understand what you are actually billed for
• AI Writing tools hub — the models these APIs sit in front of
• Best AI tools for creators 2026 — our full ranking