Kernel: Gemini 4 Carbon took close to an hour on bug-fix tasks that Opus and Astra finish in minutes
Kernel, a company that runs private coding evals, says that on bug-diagnosis-and-fix tasks Gemini 4 Carbon took close to an hour where Claude Opus 5.5 and GPT-6 Astra took a few minutes. These are Kernel's own measurements, not Google's, and it quotes users on social media who report the same slowness.
Kernel traces the gap to three causes:
- Slower inference: each step takes longer, which it says explains 16–30% of the extra time against Astra and 45–49% against Opus.
- More work: Carbon took about seven times as many steps as Astra, eleven times as many actions and processed more than sixteen times as much input. About half its steps went on work the ticket did not ask for, such as auditing code beyond the bug and guarding against cases the ticket had ruled out.
- Mechanical inefficiencies: about one tool call per step against more than two for Astra, and 39% of the lines it read were ones it had already seen, against 7% for Astra. Kernel estimates these account for roughly a quarter of the excess runtime.
Kernel says the harness is not the cause (opencode and mini-SWE-agent gave no meaningful difference), and that on low effort Carbon often failed to produce a fix. In one re-run that asked Carbon every five turns whether it had enough to make the fix, it finished in 28 minutes instead of 37.
Sources
- Why is Gemini 4 Carbon taking so long?Hacker News · 42 comments
Get news like this as a push.
AI Tracker alerts you when something changes for the tools you use.