Jev 接入 Claude Code：182 个技能选错率从 16.8% 降到 7.3%

Kravn @0xKravn原推发布于 2026.09.21

在 X 上打开原推

A former OpenAI researcher put out a model that cannot write a single sentence. about 13% of teams on Vercel's AI Gateway plugged it in within a day. it only decides. most agent stacks still pay a frontier model to write a paragraph so their code can pick one option. his name is Diogo Almeida. co-author of InstructGPT, the research ChatGPT grew out of. the company is TypeSafe, the model is Jev, in early access since September 15. with Claude Code the split is clean. Claude Code stays the slow brain that plans and writes the code. Jev takes the forks: which skill to load, which model gets the subagent. in TypeSafe's own test, a Claude agent choosing from 182 skills loaded the wrong one 16.8% of the time. with Jev picking, 7.3%. the video is the loop as I mapped it. state goes in with every question in one call, typed answers come back in 70 to 500 milliseconds. their benchmark: 13 checks on one long document took 0.27 s as one call and 2.71 s as thirteen. then your code reads the confidence. high acts, medium asks you, low goes to a human. the line almost nobody quotes: the launch post calls its own "193.6x faster, 444.6x cheaper" claim "on the higher end of real world gains." it can't reliably count and it only reads text. it's a function call, not a coworker. their launch post puts it bluntly: "Models have been superhuman at chat for years, so where is all the automation?" almost nobody shipping agents right now can tell you how many of their LLM calls were only ever a yes or no.

这条是引用着另一条写的：x.com/cyrilxbt/status/2101515130798297448