The Cheap Verdicts War: the AI Layer That Says 'Yes or No' Without Writing an Essay | inicIA

Direct answer

On October 6, 2026, OpenAI opened its Decisions API in public beta (`POST /v1/decisions`, gpt-6-luna model): an API that generates no text. It returns a typed verdict — a probability, a choice from a list, or a score against a rubric — at $0.10 per million input tokens, with no output charge (OpenAI says it is about 10x faster than its Responses API; that figure comes from the vendor, with no published independent benchmark). Its open rival, TypeSafe Jev, lists $0.042 per million tokens (published price, roughly 58% less). The message for your business: every micro-decision of an agent — approve, reject, route, qualify — does not need a frontier model writing a paragraph. It needs a cheap verdict.

What this API does (and does not do)

Three question types, per OpenAI's documentation:

It ships with Zero Data Retention and HIPAA eligibility for qualified customers, and data residency in the US and Europe. What it does not do: it does not extract fields in your shape or write explanations — for that, OpenAI still points to Structured Outputs. It is a verdicts API, not a generation API.

The math your business cares about

One million decisions of 1,000 tokens each: $100 with the Decisions API, $42 with Jev (list prices). Classifying a 500-token support message costs about $0.00005. If your agent evaluates thousands of leads, tickets or photos a day, the difference between asking a large model for a paragraph and asking this layer for a verdict is the difference between a project that scales and one that does not.

How to apply it without being an engineer

  1. Separate "deciding" from "responding": let a cheap model classify ("hot lead: 0.87") and another — or a human — act. Do not pay for prose where you only need a signal.
  2. Calibrate on your own data: probabilities are estimates. Set your thresholds (auto-approve above 0.9, escalate 0.6–0.9, human below 0.6) with labeled examples from your business.
  3. Route by cost: every call your agent makes should go to the cheapest primitive that solves that step. Any vendor quoting agents to you should show this math.
  4. Demand traceability: ask for thresholds, logs and periodic false-positive reviews. A cheap verdict without auditing is a cheap error machine.

Related reading

Frequently asked questions

What is OpenAI's Decisions API?

An endpoint in public beta since October 6, 2026 that returns typed verdicts (probability, choice or score) instead of generated text, on the gpt-6-luna model.

How much does the Decisions API cost?

$0.10 per million input tokens, with no charges for output tokens, cache reads or cache writes, per OpenAI's announcement.

What is TypeSafe Jev and what does it cost?

The open rival to the Decisions API; its published price is $0.042 per million input tokens, roughly 58% less than OpenAI's API (list-price comparison).

What is a decision model for in my business?

To classify, route, approve or qualify at scale — leads, tickets, complaints, photos — where you need a reliable, cheap signal, not prose.

Take the next step

Every decision your agent makes with an expensive model is margin you give away. Let us design your agent architecture by real cost. Talk to inicIA on WhatsApp: https://wa.me/17866120930

Chat on WhatsApp

By inicIA