gemini-4POST
5.8B-token test: Gemini 4 Argon sustains long builds but overrates itself
After two weeks and 5.8B tokens on Gemini 4 Argon, the reviewer calls it the best long-horizon builder, running 15–30 hours without drifting, yet self-scoring 9 versus 6–7 from an independent reviewer.
码农暖爸
@Delroy715Archived 14 days agoOriginally posted 2026.10.20
码农暖爸 @Delroy715Originally posted 2026.10.20
卧槽，有大佬把 Gemini 4 Argon 跑了两周、58 亿 token 后写了最诚实的评测 结论很硬：它是目前最好的长时间构建模型，能连续干 15-30 小时不跑偏，但同时也是最差的自我评价者。自己给自己打 9 分，独立 reviewer 只给 6-7 分，报告也容易 overclaim 作者最佳实践是 Argon 负责构建，配一个严格 reviewer（他用 Opus）负责检查， 做大型项目的兄弟可以看看
This post quotes another post: “Gemini 4 Argon review after 329 sessions and 5.86B tokens”
x.com/delroy715/status/2112409419829149849Archived 2026.10.21 09:54