Prism edged Jev by two emails on Arushi Gandhi's 200-email test
A one inbox, 200 item comparison of four decision models, with the winning margin small enough to be noise.
Arushi Gandhi pulled the 200 latest emails from her own inbox and asked four decision models to classify each one as cold email or not. She reports that Sarvam AI's Prism took the highest accuracy, and that the margin was two emails over Jev.
Prism: highest accuracy, but only two emails ahead of Jev.
OpenJev: most conservative - fewest legitimate emails flagged, but misses more cold mail.
Two emails out of 200 is one percentage point. On a single person's inbox, labelled by that same person, that is a tie with extra steps. Gandhi does not claim otherwise, she calls it a 200 task mini benchmark, and the more useful part of the post is the shape of the errors rather than the ranking.
What each model got wrong
Jev was the fastest of the four and had relatively few false positives, meaning it did not often shout cold email at mail that was legitimate. OpenJev went the other way. Gandhi describes it as the most conservative of the set, flagging the fewest legitimate emails while missing more actual cold mail. That is the classic precision and recall trade, and which side you want depends entirely on whether a missed cold email or a buried real one costs you more.
The fourth model in the comparison is not named in the post, and neither are the per model accuracy figures behind the ranking. There is no published latency number attached to the claim that Jev was fastest, and no cost figure for any of them.
The question nobody answered
Kush replied to the thread with the obvious follow up, asking whether Prism is cheaper. There is no answer in the thread. For a classification job that runs on every inbound email, price per call and tail latency decide the architecture long before a one point accuracy difference does, and this test measured neither.
For people building on it, treat this as one engineer's read on her own mail, not an eval you can port. The labels are personal, cold email is a judgement call that varies by who is receiving it, and 200 items is small enough that a handful of edge cases flips the ordering. If cold email triage is a product surface for you, the transferable part is the error profile, Jev leaning toward fewer false alarms and OpenJev leaning toward fewer false flags at the cost of misses. Run your own 200 and see which failure you can live with.
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.