A model that cannot write a sentence returns typed decisions for $42 per billion tokens, and its output tokens cost nothing. In its first day it reached nearly 13 percent of Vercel's paid teams, twice the GPT-5.6 family.

What a decision model answers

What it costs, against what

What it cannot do

• Text generation is out of scope. Jev gives up string generation, which rules out chat, drafting, summarising and open-ended reasoning; TypeSafe positions it as the cheap step before an expensive model call 2.
• One public head-to-head went the other way. Nikhil Mudholkar, CTO of Bryo AI, tested 1,565 German and English business emails across ten categories and found Gemini slightly more accurate, with Jev ten to twenty times cheaper. The errors' shape interested him more than the score, and he is still putting the model into production 10.
• Confidence moves work back to you. Armin Ronacher, CTO of Earendil, framed the trade-off: a 0.5 probability is a coin toss the caller has to know to discard, while 0.95 is something to act on 3.
• Encoder classifiers have existed for years. Sebastian Raschka, who trained them, reads the breakthrough as one of generalization, with the likely secret in the data rather than the training method; Kun Chen goes further, that the academic novelty is modest and the packaging produced the adoption 1112.
• The architecture, the supply behind it and the capacity are undisclosed. TypeSafe is tight-lipped about how the model is built, and outside observers suspect it sits on top of an open-weight language model; demand at launch briefly took its API down, and general availability began only on Sunday 13.

Why this week

Where your product sits

• You run an LLM classifier at high volume — support triage, moderation, intent routing, pre-filtering for retrieval. That is the call to price against $42 per billion input tokens 2.
• You are building guardrails for an agent. LangChain's integration ships a middleware that checks a tool call before it executes, and another that picks a model per request 7.
• You ship a free or ad-supported consumer app. Andrew Chen of a16z argues the economics of a per-screen model call finally work at a thousandth of the old price, which is what a free AI-native app needs 15.
• You buy AI for an enterprise workflow. Aaron Levie of Box demonstrated incident reports pulled from a content system, classified, sorted into escalate, monitor or review, and tagged with metadata — the shape he expects in claims, contracts, loan processing and security review 16.
• You are the platform selling the loop. A Vercel engineer replaced the GPT-5.6 Luna classifier behind a safety feature with Jev and reported five to eighteen times the speed at higher accuracy, and builder demos in the first week covered browser use at roughly a tenth of a cent per task 171819.

A pilot you can defend in a review

1. Pick one decision, not a category. The public tests that produced usable results each chose a single question — is this urgent, which department, is this action safe — and one labelled corpus to answer it against 1020.
2. Assemble the labels before the integration, then measure agreement with the call you make today. The two tests worth reading sampled 1,565 business emails and 1,500 personal emails, checked by hand, and logged where the model and the human call disagreed rather than how often 1020.
3. Write the escalation band before you run the pilot. Decide which probability sends a case to a person, then measure how often that band fills up 3.
4. Keep the seam explicit, with the decision on the cheap model, the open-ended work on the LLM you already pay for, and the two meeting in one function your team owns 7.
5. Set the cost gate on your own traffic, not against the vendor's 444-times figure, which came from workflows built in-house 2.

What changes when a decision model enters the loop

Evidence boundary: Every cost, latency and accuracy figure here is TypeSafe's own, from evaluations the company built and published; the evals take the average of a GPT model and a Claude model as the reference answer, which the company's own notes say biases the comparison toward those two labs, and it states it cannot prove its pricing is not subsidized. The adoption number is Vercel's telemetry for its own paying teams, which measures uptake inside one gateway rather than market share. The counter-evidence is a single startup CTO's benchmark of 1,565 emails and two practitioners posting on X, and nobody has replicated the headline claim independently.