gemini-4IMAGE

FrontierCS 2 基准：Gemini 4 Argon 40%，落后 Opus 5.5 的 58%

新基准 FrontierCS 2 要求 agent 在看不到论文的情况下改对真实 CS 研究代码，Opus 5.5 通过率 58%，GPT-6 Astra 44%，Gemini 4 Argon 40%，换 harness 后 Argon 接近 Opus。

vorpal

@vorpal_onchain4 天前收录原推发布于 2026.10.26

FrontierCS 2 基准：Gemini 4 Argon 40%，落后 Opus 5.5 的 58%

vorpal @vorpal_onchain原推发布于 2026.10.26

在 X 上打开原推

Opus 5.5 is still the best researcher on earth the new FrontierCS 2 benchmark makes agents beat real CS research code without seeing the paper Opus 5.5 leads the new FrontierCS 2 benchmark Opus 5.5 passed 58% of runs. no other agent reached 50%, Astra 44% and Gemini 4 Argon 40% swap Argon into Antigravity and it almost catches Opus the harness matters as much as the model Terra has a lot to prove this week is your harness holding your model back?

x.com/vorpal_onchain/status/2114634678359457938采集于 2026.10.27 09:52