The Decision Layer: Stop Paying Frontier Prices for Yes-or-No Answers

In September 2026 a startup called TypeSafe released Jev, a model that never writes a word. It reads a piece of text and a set of typed questions, and returns typed answers: which option, with a probability for each; a score on a scale you define; the probability that a statement is true. It costs $0.042 per million input tokens, and output is free. Three weeks later OpenAI opened a Decisions API that works the same way at $0.10. A day after that, Anthropic released Claude Haiku 5.5 and described it as built for classification, routing and extraction. On October 9, TypeSafe raised $870 million at a $7.5 billion valuation.

Most companies make their AI's routine judgments, which queue, which category, is this a duplicate, does this source support this claim, on a general model built to write. At list prices, one million such decisions cost about $1,900 on a mid-tier model like Claude Sonnet 5.5 or OpenAI's GPT-6.1 Sol, $95 on Haiku 5.5, $70 on OpenAI's Decisions API, and $29 on Jev. A decision model also answers in a type your code can use without parsing, in a tenth to a third of a second, with a probability your code can route on.

How much of your AI work fits? No one publishes the figure, so we measured the closest thing to it. In Anthropic's public sample of its business API traffic for May 2026, about one conversation in six is a decision job: tagging, scoring, screening, checking, routing or extracting into a fixed structure. Counts of automation workflows put the share higher, at more than a quarter. By tokens and spend the share is smaller, because coding and long conversations dominate both.

The evidence behind the new models is thinner than the marketing. On one independent benchmark, Jev's accuracy rose from 62.6% to 95.0% when a single question was split into five narrow ones, and a two-line rule scored 91.8%. No vendor publishes a calibration measurement. Question design, not model choice, is most of the work.

This paper is the method. It explains what decision models are and are not, compares every option on price, limits and data controls, and gives a cookbook of fourteen decisions worth moving, from ticket triage and model routing to citation checks, extraction verification and agent tool choice, each with example questions and thresholds. It sets out ten rules for writing questions, a two-week evaluation protocol with an acceptance test written before the run, five architectures that use the confidence, and the break-even arithmetic for adding a vendor such as TypeSafe instead of using the decision option your current provider already sells. It closes with a way to measure the return and a ninety day plan by owner. The decision inventory, the evaluation scorecard and the question design checklist are included as templates.

The argument in one line: move your AI's decisions to the cheapest model that passes your own test, and spend what you save on the checks you could never afford.