WHAT NEW IN X10.28 WED

AI collections from X · 22 collections · 58,925 posts

New todaydeepseek-v4.1IMAGE

Cognition's FrontierCode 2.0: DeepSeek V4.1 Flash 19.3% vs Opus 5.5 45.5%

The new coding benchmark from 20+ maintainers leaves requirements implicit and zeroes hard-rule violations; DeepSeek V4.1 Flash scores 19.3% against Opus 5.5's 45.5%.

墨无归
@MoblacksoilArchived 10 hours agoOriginally posted 2026.10.28
Cognition's FrontierCode 2.0: DeepSeek V4.1 Flash 19.3% vs Opus 5.5 45.5%
墨无归 @MoblacksoilOriginally posted 2026.10.28
Open original on X

做 Devin 的 Cognition 发布了编程评测 FrontierCode 2.0:题目由 20 多位开源维护者从自家项目出,每题花 40 小时以上,而且故意不写全,看 AI 能不能像靠谱同事一样,补上没说出口的要求。 上一版问「这个改动会不会合进项目」,这版再问「该不该合」。 分数是按评分细则加权的总分,踩到硬性红线直接记 0。最高的 Claude Opus 5.5 只有 45.5%,开源的 DeepSeek V4.1 Flash 是 19.3%。这套题短期内还很难被刷满。 https://x.com/cognition/status/2115132488775524793/photo/1

x.com/moblacksoil/status/2115181994262839493Archived 2026.10.28 12:04