Google's Gemini 3.5 Flash and Anti-gravity CLI Critique
Theo - t3.gggo watch the original →
the gist
Google's new Gemini 3.5 Flash model and closed-source Anti-gravity CLI suffer from poor token efficiency, high costs, and buggy performance, while internal politics have sidelined the team behind the previously successful open-source Gemini CLI.
Performance and Cost Inefficiency
Gemini 3.5 Flash demonstrates high benchmark scores in agentic tasks but fails significantly in practical application. Despite marketing claims of speed, the model exhibits poor token efficiency compared to competitors like GPT-55 Medium. The cost structure has effectively tripled, with pricing now at $1.50 per million tokens input and $9 per million tokens output. In real-world testing, the model produced broken, non-functional code for a game-development task, requiring multiple iterations that resulted in further regressions, whereas other models completed the task successfully on the first attempt.
CLI Regression and Internal Politics
Google has deprecated the open-source Gemini CLI in favor of the new, closed-source Anti-gravity CLI. The new tool is written in Go, suffers from significant UI bugs such as broken scrolling and persistent input field issues, and lacks the community-driven development that defined its predecessor. The transition appears to be the result of internal corporate politics, where the original team—who maintained strong community relations and open-source progress—was sidelined in favor of leadership brought in from external acquisitions. This shift has resulted in a product that mimics existing tools like Codeex while failing to match their functional stability.