Jev 浏览器操作与即时评估实测

Aakash Gupta @aakashgupta原推发布于 2026.09.26

在 X 上打开原推

This guy helped create ChatGPT. His new startup is in talks at a $10 billion valuation. Meet Diogo Almeida. He worked at Google Brain. He worked at OpenAI. He was a co-author on the RLHF training paper behind ChatGPT. Then he left. He spent two years in stealth with his co-founders, Erik Gafni and Sasha Sheng, on a $40 million seed led by DCVC at a $200 million valuation. On September 15, they finally came out of stealth and launched their first model, Jev. The launch video has 38 million+ views on X. And Jev became the fastest-adopted model in the history of Vercel's AI Gateway, with 13% of paid teams using it in the first 24 hours. That's both viral success and adoption, which is pretty insane in an AI market that's incredibly crowded and has pretty much been dominated by two players for much of the last three years. We've seen lots of new model launches from big names, including Mira Murati. But Diogo Almeida seems to have shaken up AI the most with Jev. I spent the last week with it, and I haven't been able to stop using it. Here are just a couple use cases that are pretty interesting. First, it can do browser use much more cheaply. We all know browser use is one of the most expensive tasks we're doing with models like Opus and Fable right now. Fable charges $10 per million input tokens and $50 per million output tokens. Jev charges $0.042 per million input tokens, and doesn't even charge for output. So it can very, very cheaply use browsers, and it's actually quite accurate. It doesn't hallucinate anything because of the architecture it's built on. It can only pick from the answers you define. Another interesting use case is doing any sort of instant evals, which is really important for product teams. You can instantly get live evaluations of things at almost no cost. In fact, in the last week I've been using Jev, I haven't even blown past the $5 in free credits they start you with. That's how cheap it is. And it's basically instantaneous. It's much faster than an LLM because an LLM writes its answer one token at a time, and every token waits on the one before it. Jev doesn't write anything. It answers every question you give it in parallel and hands back a decision with a probability attached, in 70 to 500 milliseconds.