Gemini 4 Argon agent test: 142 requests and millions of tokens burned
On a simple task, Argon made 142 requests, compacted context twice, exceeded 800K context and burned millions of tokens, which the author calls terrible work habits.
This is the most painful Gemini 4 Argon test I've seen so far. The task was simple but Argon made 142 requests, compacted its context twice, hit 800K+ context and burned millions of tokens. GPT-5.6 Sol could've done the same task in under 100K. Where did it all go? Reading files that had nothing to do with the task, and testing and checking stuff that are not needed. And when told three times to stop and report progress, it just kept going. Maybe Argon isn't dumb. but its work habits are just terrible, and you pay for every one of them.
This post quotes another post: “Hands-on: Gemini 4 Argon is good at tools and reviews, not complex system code”