The First Week of Jev: A Launch Told Through What Got Built

Model launches usually follow a script: announcement thread, benchmark chart, three days of takes, quiet. Jev's mid-September 2026 launch broke the script in one specific way: the takes got outrun by the builds. Before the discourse settled on whether a decision model was a big deal, working demos were answering the question empirically, in public, with receipts. This site exists because that flood needed a catalog; this post is the launch story told the only way we trust: through what actually shipped.

We won't pretend to a day-by-day chronicle we didn't keep. What we have is better: 600+ entries, each timestamped by its author's own announcement, forming a pretty legible arc.

Act one: the speed proofs

The first wave was builders doing what builders do with an extraordinary claim: trying to break it in public. TypeSafe said instant structured decisions; the community's response was to wire the model into places where slowness is visible as failure. Doom at roughly ten rulings a second, Mario on reaction timing, Slay the Spire 2 at 0.7 seconds a move: the games category filled up not because anyone needed an AI playing Doom, but because a livestream can't lie about latency. The speed claim survived its stress test as entertainment, which is the most credible way any claim survives anything.

Act two: the economics proofs

Wave two asked the other question: is it really that cheap at workload scale? The answers arrived as the numbers this site now quotes constantly: 500 emails classified for 3.5 cents, 26,000 judge verdicts for 13 cents, browser tasks around a tenth of a cent, 700 leads scored in 40 seconds for 9 cents. Author-reported, independently convergent, and directionally unambiguous. Somewhere in this stretch the community's question audibly shifted from "is it fast and cheap?" to "what does the world look like if decisions are free?", which is the question the rest of the catalog answers.

Act three: the architecture emerges

The most consequential wave was the least flashy. Builders stopped demoing the model alone and started composing it: the fraud pipeline pairing Jev with a frontier model for a reported 96/100 at ~$0.07, verification layers judging agents' claims, confidence gates routing hard cases upward. By week's end the cascade wasn't a clever build, it was the default shape, converged on independently by people who'd never met. That's how you can tell an architecture is real: nobody had to evangelize it.

Act four: the weird stuff

And then the entries that made us keep the "Tools & Apps" category biggest of all: a Postgres function putting judgments inside WHERE clauses, instant context compaction for coding agents, semantic word games, screenshot-free computer use, a $10,000 live-trading experiment we catalog with a raised eyebrow (the trading post explains the eyebrow). Week one of any platform is when the improbable stuff gets tried, because nobody's told anyone what it's for yet. We consider it our job to make sure that window stays documented.

What launch week actually established

Not that Jev is "better", the limitations are real and cataloged, and no official benchmarks exist yet. What week one established, with unusual rigor for a launch, is what kind of thing this is: a decisions-everywhere primitive whose speed and cost claims survived public falsification, whose best-practice architecture emerged by convergence, and whose ceiling is currently set by builders' question-writing craft more than by the model. Week two's story is already being written in the submissions queue. We'll keep the receipts.

Frequently asked questions

When did Jev launch?

Mid-September 2026, via TypeSafe AI's announcement on X, with the demo flood following within days. This directory launched the same week to catalog it.

What was the first big Jev demo?

No single canonical first; the earliest wave clustered around real-time games and rapid classification runs, which is why those entries anchor the games and triage categories.

How many things have been built with Jev so far?

600+ cataloged here at the time of writing, across eight categories, each with author-reported numbers. The use-cases tour is the guided walk.

Where can I follow what gets built next?

The directory front page stays current, and shipping something yourself remains the best seat: submit it with your numbers and join the timeline.