Curated by AY Automate
21 public builds, tests and guides from X and LinkedIn in the first days after TypeSafe launched Jev on September 15, 2026, sorted by use case. Numbers are each poster's own claims. We include the tests where Jev fell short.
A demo where the open-source Browser Use agent uses Jev for its per-step decisions and finishes a one-way Google Flights search on its own.
7.1 s end to end, 150 to 300 ms median decision (poster's figures)
Masahiro Chaen on X
Side-by-side run of Jev and DeepSeek sorting 300 mixed emails (procurement, outages, legal, spam) into 15 departments with an intent label.
9.9 s and $0.036 for all 300 (poster's figures)
SUOHA_AI on X
A split workflow on a 500-item queue. Jev does the reading and sorting, and a large model only writes the 25 replies that need writing.
$0.02 a day for reading vs $0.73 for the replies (poster's figures)
A racing game where your car moves further the bigger the semantic gap between two words. Jev scores each pair.
Marcel Pociot on X
Notes on the Ably and Vercel demo where Jev made many Pong paddle decisions while chat models were still on their first move.
about 200 ms per decision (poster's figure)
Sam Parker on X
A drone autopilot that flies from point A to point B and avoids obstacles, built with the Vercel AI SDK, with code on GitHub.
$0.01 per trip (poster's figure)
A video of Jev playing Doom in real time. The poster calls it accurate at logic-based and probabilistic decisions but not creative ones, and a cheap way to simulate players.
An architecture sketch on Cloudflare: Workers for data, Workers AI with Jev for the decision, Workflows and D1 for the confidence gate and approval queue, Durable Objects for positions.
Miyata on X
A long-form walkthrough of building a trading loop that makes one buy or sell decision per block.
under 100 ms per decision (poster's figure)
Roan on X
Jev is available as an evaluation model in Braintrust. The poster says swapping the model in an existing scorer cuts scoring cost by 400x.
Jev answered two questions per tweet (what is this about, is it widespread) across seven brands. It flagged more outages and earlier than a keyword list, though the poster says the difference against a good keyword list was not statistically significant.
$1.84 for all tweets, 37 of 40 spot checks correct (poster's figures)
Jason Zhu on X
A local web app built in three prompts that combines Jev with map data, voice narration and a short generated demo video.
Franci Penov on X
Jev answers design questions (font, radius, spacing) as probabilities to build UI components from a description.
Early prototype with the goose agent harness: Jev reads the prompt and selects which model should run that turn.
A practical thread: put Jev in front of your model, ask five small questions instead of one big one, never ask Jev whether it can answer, and run it quietly beside your current system for a week.
Ronin on X
Steps for calling the decisions endpoint through OpenRouter while the direct waitlist was still open, plus a Tetris demo.
$0.042 per million input tokens, output free
Ziwen on X
Explains the shift from next-token prediction to a model that returns weighted scores and probabilities, with examples like judging whether a new prospect is a good fit.
Says structured output makes Jev a fit for classification jobs like model routing, tool selection and search, and guardrails, and that it is fast and cheap next to LLMs on the same tasks.
A matched-case test from loan origination. Jev answered "income not verified" on all twelve, and the poster says six of those were false because the transcript was partial. Giving it the full event log kept accuracy at 50% and raised its confidence.
A short caution that moving workloads off frontier models onto Jev wholesale is a questionable AI practice.
Runs Jev on a real workload, 600 labeled emails against a classifier trained on the poster's own inbox, instead of a demo, after questioning the launch statistics of 40x to 200x faster.
Jev is TypeSafe's System One model. It does not write text. You send it a state and typed questions, and it returns a choice, a yes/no probability or a score, each with a probability and a confidence value.
Since the September 15, 2026 launch, public posts show browser agents, email and ticket triage, real-time game and drone control, trading loops with confidence gates, evaluation scorers and research labeling. This page lists them by category.
No. Every figure is the poster's own claim, marked as such. AY Automate has not rerun these projects. For our own test, see the independent Jev benchmark on our blog.
Not yet. We index what the community is building and test where Jev fits. If you have a routing, triage or scoring workflow you want to pilot, book a call.
Post it publicly on X or LinkedIn, then send us the link through the contact page. We list builds that show what was built and how it performed, including the ones where it did not work.
Have a build to add? Post it publicly, then send us the link.