Theo and Diogo Almeida agree: Jev demos oversell it
Theo says he is tired of calling out bad Jev ideas, and Diogo Almeida agreed that outlier demos lead to over-promise and under-deliver. Meanwhile elvis got Jev Router to half the cost of a GPT-6 Sol baseline on eight support cases, the opposite of what Theo's code benchmark found.
Built with it
David shipped jevgrep, says it cuts coding agent cost 40%
David released jevgrep, a research agent CLI that uses Jev to gather context before the coding agent spends tokens on it. He says it reduces coding agent cost by 40%, verified on SWE-bench, and ships a skill so the agent knows to call jg for context collection.
Someone measured it
elvis: Jev Router under half the cost of GPT-6 Sol on 8 support cases
elvis ran Jev Router against a fixed GPT-6 Sol baseline on a small support agent built with the Pi SDK, 8 cases and 32 real calls, and both got every case right. The router cost $0.008 against $0.018 and had a median response time of 1.5s against 1.9s.
One of my biggest concerns about Jev-as-a-Router is caching.
The case against
Theo and Diogo Almeida agree the Jev demos are overselling it
Theo said he is tired of watching people push Jev at problems it is not good for, and Diogo Almeida agreed. Theo's line on where it does work: classification of fixed data with mostly known quantities, which he calls System One, as opposed to models with System Two reasoning.
Posting outlier demos leads to over-promise under-deliver