Ten days after launch, Jev — TypeSafe's blazing-fast 'system one' judgment model — has gone from viral emoji demos to real production use. Today: six categories of actual use cases, from ad teardowns and live-prioritized inboxes to agent routing, plus where the model breaks down and how to write questions it can actually answer.
Cheap, instant judgment at scale is a new primitive — not another chatbot.
Jev can't write code or draft contracts. What it can do is make small judgments — which bucket, how urgent, is it relevant — in milliseconds, for fractions of a cent, across everything you have. Ten days in, the pattern is clear: when judgment gets 40 to 400 times cheaper, you stop sampling and start asking questions about every ad, every email, every page, every agent step. That's not a difference in scale from LLM review — it's a difference in kind, and it's going to take a while to figure out how deeply it weaves into everything.
The Information reports TypeSafe is in talks to raise as much as $1 billion at a $10 billion or higher valuation — a huge jump from its $40 million seed at $200 million. Hundreds of viral posts have translated directly into financial opportunity.
TypeSafe borrows the term from Kahneman's Thinking Fast and Slow: fast, instinctive pattern matching. The test for a good Jev job: anywhere you repeatedly read something, make a small judgment, and take a predictable next step — like deciding a downloaded PDF is an invoice, which project it belongs to, and who needs to see it.
Choice: pick the best fit from up to 255 options, with probabilities for every option. Score: place something on a 2-to-10-level scale you describe in words. Null (short for Bernoulli): a yes/no with a 0-to-1 probability that the statement is true. That's the entire interface.
Stefan's visualization: type 'things you can wear in winter' and matching emojis race to the top. Every emoji gets the same yes/no question live — nobody labeled or tagged anything, and there's no rule for 'winter.' Because what Jev does is limited and specific, it can do it incredibly fast.
TypeSafe's launch claims versus comparable LLM processes. A million input tokens cost 4.2 cents, output is free, and each call takes 70 to 500 milliseconds. The power isn't just that it's good at judgment — it's that it's fast and cheap enough to make judgments about everything.
Questions run in parallel, so asking more doesn't cost time, just tokens. TypeSafe found 13 questions in a single call was 12.2× cheaper and 10× faster than asking one at a time — with identical answers.
Matthew Berman ran 12 questions per ad — hook archetype, format, offer, CTA intent, awareness — across 724 live ads from 37 brands; each 12-question call took just 173 milliseconds. Compare that structural data to actual performance and you get fine-grained advertising insight that was previously impractical to produce.
Berman then had Jev evaluate 723 ads as 30 buyer archetypes — gym owner, dental office manager, toddler mom — asking for every pair: would this person stop scrolling? These aren't real buyers, so treat it as a hypothesis generator. But it's hard not to think marketing will use this to pressure-test ad and landing-page angles before paying for real tests.
Ian Nuttall ran eight questions — topic, hook, tone — across nearly 3,300 of his past X posts for about 13 cents, then compared against engagement to see what actually worked. Any big pile you have — a year of customer emails, CRM notes, a folder of transcripts — is a potential Jev job if you can write a few simple questions.
Justine Moore of a16z described scanning thousands of Zillow listings for things you can't filter for — architectural style, renovation status, freeway proximity. Burhan clipped a 90-minute video by natural-language themes like 'their predictions for when AI will automate AI research' — in under two seconds, for under two cents.
Borja had Jev read all 586 pages of a website and rebuild the internal link map — 8,790 yes/no calls — in 45.1 seconds for 21 cents. Claude Opus 5 got through 21 pages and spent $1.43. As Borja put it, internal linking is a classification problem 'and we have been paying frontier prices to do it one page at a time.'
Robin built a live AI-slop detector that confidence-scores every X post as you scroll and blocks past a threshold. Elvis had Jev read 384 morning news stories and flag which ones 15 different brands should care about — 24.9 seconds, 19 cents. In the same window, Opus 5 got through four articles.
Jonathan Yunikowski asked: what if your inbox was live-prioritized by importance instead of reverse chronological? Jev rated 100 emails in 453 milliseconds for about a tenth of a cent — and matched his own ratings on every single one. 'This is a use case that I want right now, not tomorrow. In fact, I wanted it yesterday.'
Marcel Pochulu built a macOS app that watches his downloads folder — is this an invoice? file it correctly — with no other LLM calls. DevEd runs live chat moderation. Stephen Tey of Dub fed Jev 10,000 previously caught malicious domains to flag bad links: 'something we've been wrestling with since day one. With Jev, we solved it in two hours.'
Most workflows eventually hit the same question: what should happen next? Box has experimented with incident triage — customer impact, severity, escalation routes — and Odoo CRM has a proposed module asking three things about every lead: priority, buying readiness, spam or not. If you're in a company with multiple people touching the same leads or tickets, this is the one area to explore first.
Harrison Chase of LangChain called Jev 'great for evals, especially online evals where you want to grade lots of traces.' That sounds dev-only, but as everyone starts putting support bots and research agents into production, a cheap tool for continuously judging agent work looks like core infrastructure, not a niche.
Every planted mistakes in 12 passages: Jev caught six of seven versus Claude Fable 5.1's seven — but in 0.35 seconds versus 8.83, at roughly 580× cheaper. At that price you can rerun the same check enormously many times and still beat a single frontier pass. Practical version: turn your style guide into yes/no questions and run them on every paragraph.
The biggest rewiring Jev demands: its value isn't just doing a thing, it's doing it at such scale that the capability changes category. Checking every single sentence in minute detail for AI-isms is categorically different from running one generic LLM review across a whole document.
People are using Jev to route agents to the right model, reasoning level, and tools. Vi Chen had Jev adjust GPT-6's reasoning effort inside Codex mid-task — more thinking when stuck, less for routine steps — cutting cost about 50% while runs got faster.
Daniel Son's Jev Skill Suggestion for Claude Code classifies which skill best matches each request and injects only that skill into context — an 88% decrease in tokens and cost. If you've built up a big skills library in Codex or Claude Code, this is how you stop shipping the whole thing with every request.
AJ Asver's team built a Jev-based harness that cuts repetitive-work costs by 90%, learning as it runs and moving steps from LLM calls to code. On compliance alerts, cost per alert fell from about $2.95 at the start to 25 cents by alert 1,000. Expect judgment models to be built natively into the tools and harnesses we use.
Marcus Lowe's demo: copy a resume, paste into an application, and the fields fill themselves. Norman's version splits pasted text, works out what each field is, and pastes only confident matches. Given how much work is moving details from emails, PDFs, and notes into CRMs and forms, this category feels very, very relevant.
Ranking job applications with Jev is tempting, but even with good criteria you get a number and no rationale. The better pattern is hybrid systems: let Jev rank at volume, then automatically send bubble candidates to a fuller LLM review for a second look.
Multi-step questions, where accuracy drops with each hop. Counting, math, and dates — it can extract the facts, but do the arithmetic elsewhere. Plus known issues with consistency and reading intent. Refreshingly, the vendor is explicit about the limits.
One: you can write the possible answers in advance — categories, a yes/no, a scale. Two: there's a pile or a stream, hundreds of tickets or emails. Three: a wrong answer is cheap or easy to catch — don't leave high-stakes calls to its judgment alone. Four: you can hand it the evidence as text in under 32,000 tokens.
Break questions into constituent parts: industry fit, company size, buying intent, and so on, each as its own question. Describe every scale level in words, ask more questions than you think you need, and test before you trust — label 50 emails yourself and compare before letting the classifier run your inbox.