两周 329 场会话后的 Gemini 4 Argon 实测长评
作者用 Argon 跑了 329 场会话、58.6 亿 token、404 小时自主工作，翻遍日志写成一份优缺点俱全的诚实评测。
So let's talk about Gemini 4 Argon.. I've been using it for around ~2weeks. 329 sessions. 5.86 billion tokens. 404 hours of Argon working on its own. I went through every single log to write the most honest review I could.. the good, the weird and what it still needs to fix 🧵
Now here's what i honestly want! See it's a pretty big step up from the flash series.. some of the things and problems i have mentioned are project specific or might just have originated from how i work so performance & results can be different.. GDM team did a good job with Argon but its not just the model its the harness too.. I don’t know why agy caps it at ~256K when the model has a 1M window I hope those enforced limits can be removed so that we can truly see how good the model is.. Compaction is pretty bad.. like the model forgets the facts of the project(sometimes) on long running tasks.. pls make the compaction and cache hit rates better(after compaction) I expected a faster Gemini. What I got is a model that holds a long, complicated task together better than anything I've used, and that can't yet be trusted to tell you how well it did. Pair it with a strict reviewer and it's a very strong builder.