Jev Daily

Jason Cohen says Jev could not beat Sonnet on classification

A public ask for anyone still running Jev in production, and one reply from someone who is.

Jason Cohen asked publicly whether anyone is actually using Jev, and said his own attempt did not go well. He tried it on what he describes as many classification jobs and found it hard to get it just to match Sonnet. That is the whole of the negative result, stated by a person who ran it rather than someone reading about it.

Jason Cohen
@asmartbear
X
I tried it for many classification jobs and it was hard to get it just to match Sonnet.
Nov 16, 2026 · View on X

The second half of his post is a signal about attention rather than quality. He says that a month ago the hype was amazing, that a YouTube search turns up videos from then, and that nothing newer comes back. That is an observation about one search on one platform, not a usage number, but it is the kind of thing a founder notices before deciding how much of a product to hang on a dependency.

Isaac Flath replied that he uses Jev for many things, both personal workflows and commercial. That is one counterexample from one account, with no task or numbers attached either.

What is missing

Nothing in Cohen's post is reproducible. He does not name the classification tasks, say how many classes were involved, describe what his accuracy criteria looked like, or give the cost or latency he was comparing against on Sonnet. There is no mention of how he structured the calls or how many examples he gave. Match Sonnet is a verdict, not a measurement, and without the task definition a reader cannot tell whether the gap was in the model or in the harness around it.

For people building on it

Treat this as what it is. One experienced builder tried Jev on a class of task it is routinely recommended for, could not get past his existing baseline, and said so in public. That is worth more than a vendor claim and less than a published eval. If you are choosing between Jev and a general model for classification, the useful move is to run your own labelled set against both and record the accuracy and the per call cost, because neither Cohen nor Flath has published anything you can copy.

Isaac Flath
@isaac_flath
X
Yeah, I use it for many things. Both personal workflows and commercial.
Nov 17, 2026 · View on X

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.