Jev AI Is INSANE... 193x Faster Than LLMs?!

Play the video (opens in a new tab)

Rithesh walks through TypeSafe’s launch post and reads the fine print along with the headline numbers. The side-by-side demo against GPT-5.6 Terra shows every field in Jev’s answer filling in at once while Terra works down the list. The two disagreed on one field, and TypeSafe admits the demo query was simplified. On the workflow evals, Jev scores about 68% accuracy at a small fraction of the cost. That is where the 193.6x faster and 444.6x cheaper figures come from, and TypeSafe calls them the high end.

The video points out several catches. Jev is not the most accurate model on that chart; two larger models score in the low 70s. Accuracy there means agreeing with GPT-6 Astra and Fable 5.1, and TypeSafe’s own team wrote the workflows. The 0% type-error bar was not measured: Jev’s output always fits the schema by design, so it cannot return the wrong type but can still return the wrong answer.

The demos get the same treatment. The Doom bot makes about 10 decisions a second for roughly $7 an hour, reads the game as text rather than pixels, and loses to a hand-coded bot. In Wiki Racing the LLMs ran with little or no reasoning. The video also cites an early tester who ran about 5,000 requests for about $2 at around 150 ms per call, and another who classified 1,500 of their own emails.

Open on YouTube