Why is Gemini 4 Carbon taking so long?

271 points · 139 comments on HN · read original →

Points and comments are a snapshot, not live.

Gemini 4 Carbon is up to 4x slower than rivals due to unnecessary work and mechanical inefficiencies.

Kernel.dev benchmarks show Gemini 4 Carbon takes ~22 minutes on bug-fixing tasks that Opus 5.5 completes in ~9 minutes. Carbon's slowdown has three sources: slower inference (16-30% of gap vs Astra), unnecessary work (half of its steps audit code beyond the fix or close loopholes already ruled out), and mechanical inefficiencies (fewer tool calls per step, 39% of lines re-read). When prompted every 5 turns to check if it has enough info, Carbon's wait time before editing shrank from 25 to 10 minutes, suggesting the behavior is habitual rather than a knowledge gap.

What commenters are saying

Commenters split into two camps: those frustrated by Carbon's slowness and those praising its output quality. Several note Carbon's cautious, thorough approach produces less overengineered code than alternatives. Some suggest Google prioritized benchmark performance over efficiency, with one commenter speculating Carbon may be ~10x the size of Opus 5.5. A few recommend using Carbon for open-ended tasks and Hydrogen for speed-critical work, noting Google's own blog suggests using both models together.