I truly didn't understand the value of Jev. I read the launch posts, I read the docs, and I still had a hard time seeing how it was unique. So I built a test to see it in action.
Quick background. Jev is a new model from TypeSafe AI, and it doesn't write anything. No chat, no code, no summaries. You hand it some text and a question with a fixed set of answers, and it hands back its pick and how confident it is. That's the whole product. On paper that sounded to me like a classifier with good marketing.
The race
I generated 20,000 emails where I already knew the right bin for every one: urgent, client, newsletter, or junk. About 30% are built as traps. A newsletter with URGENT in the subject line. A calm little "quick one" note where the customer walks this afternoon. A fake invoice from a lookalike domain. Then three sorters ran the same pile at the same time. Jev, Claude Haiku 4.5, and a plain set of keyword rules. Same harness for both models, 50 emails per call, 16 calls in flight.
| Sorter | Time | Correct | Cost |
|---|---|---|---|
| Jev | 4.6 seconds | 99.7% | $0.15 |
| Haiku 4.5 | 76 seconds | 99.9% | $1.91 |
| Keyword rules | instant | 78.1% | $0 |
Haiku was more accurate. I want to say that plainly, 99.9% beats 99.7%, and Jev put 56 emails in the wrong bin. But Jev was done before Haiku was a tenth of the way through, about 16 times faster and 12 times cheaper. And the keyword rules, the free instant option, got more than one in five wrong. A newsletter that says URGENT goes straight to the urgent pile.
Jev is in its own class. It's a different economic model.
Why that changes what you'd build
That bottom row is how most software makes decisions today. Rules and filters, because real judgment has been too slow and too expensive to run on everything. So we sample. We audit once a quarter. We make people fill in dropdowns.
At these numbers a million emails is about $8 and under four minutes. That's where I think a lot of new use cases open up. Cybersecurity could analyze every event instead of leaning on rules and filters. A CRM doesn't have to keep asking people to enter and update dropdown fields, it can read the email and the call notes and set them. Insurance and accounting could become continuous audit, while the work happens in the software, instead of a review after the fact.
What I haven't proven
This is one run on emails I made up. Real inboxes are messier than my traps. And the part I care about most is still open. Jev gives a confidence score with every answer, and I don't know yet whether the 56 it got wrong were the ones it was unsure about. If they were, you send those few to a bigger model or a person and let the rest fly. If they weren't, it's just a fast classifier that's slightly less accurate.
So next it runs in shadow on my real inbox. It gates nothing, it just logs its answer and its confidence next to what actually happened. I'll know in a week or so.
