Gemini 4 Argon 高强度推理实测：3D 游戏与额度吐槽

Kavish Krishnakumar (e/acc) @SynKavish原推发布于 2026.10.14

在 X 上打开原推

A few more observations after spending more time testing Google Gemini 4 Argon on High reasoning: 1. It is a bit odd that Google still only provides low, medium, and high reasoning efforts, without an xhigh or max thinking mode like OpenAI and Anthropic offer. I suspect this might be because Gemini models have an innate tendency to overthink, or worse, doubt themselves and spiral into a lot of self-hatred. I hope Google has fixed this, though one can only hope since we can no longer see the CoT (chain of thought). 2. This model has a great understanding of dimensionality and the world, though it isn't the best at one-shotting 3D worlds. I have been testing a London-Bombay Zombie FPS test game called 'Mind the Gap'. One-shotting this game with Argon gave me poor results at first. However, having Claude Opus 5.5 act as an adversarial reviewer for the game models and GPT-6.1 Sol as an adversarial reviewer for the logic - and then handing those reviews back to Gemini - gave me some incredibly impressive results. I think part of this is to do with the major labs recently receiving a large training data corpus on 3D stuff and geometry, while another part is that Gemini still needs some work in the arena of 'taste' - which currently Opus 5.5 is king at. The model is genuinely intelligent, it just needs a lot of steering in order to work with it. Also, the usage limits are just absolutely sordid right now, and the fine folks at DeepMind should probably seriously consider raising limits for the Pro plan and above to have a real chance at competing at the big boy level.