An agent that keeps investigating
Gemini 4 Carbon took close to an hour on software-debugging tasks that competing systems completed in minutes, according to a November 18 analysis published by Kernel. The evaluator traced the models step by step and concluded that Carbon's delay came largely from doing far more work than the task required, rather than solely from slow token generation or network latency.
The private evaluation gave each agent a codebase captured before a known bug was fixed and a report describing only the symptoms. The agent had to identify the cause, make a targeted repair and avoid unrelated changes. That setup tests not just code knowledge but judgment: finding the relevant path, rejecting alternative explanations and recognizing when enough evidence exists to act.
Kernel reported that slower inference explained between 16% and 30% of Carbon's added time compared with Astra, and between 45% and 49% compared with Opus. The remaining gap came from longer action sequences. Against Astra, Carbon averaged about seven times as many steps and eleven times as many actions such as reading a file or executing a command. Because the growing conversation was sent again at each step, it processed more than sixteen times as much input.
Repetition and delayed commitment
The analysis found that about half of Carbon's steps addressed work outside the requested ticket. In one illustrated run, both models located the likely cause early. Astra wrote a fix several minutes later, while Carbon continued investigating for roughly another 25 minutes before editing code.
A modified trial added the question, "do you already have enough to make the fix?" every five turns. In that test, Carbon finished in 28 minutes instead of 37, and the interval between identifying the cause and making an edit fell from about 25 minutes to roughly 10. Kernel interpreted that result as evidence that the model's operating behavior, not a missing piece of technical knowledge, was a significant source of delay.
Mechanical choices compounded the problem. Astra often issued multiple tool calls in one step, while Carbon generally made one call and read one file. Carbon also revisited code much more often: 39% of the lines it read had already appeared earlier in the run, compared with 7% for Astra. Kernel estimated that limited batching and repeated reading accounted for around a quarter of Carbon's excess runtime.
The findings come from a private evaluation rather than a broad public benchmark, so they should not be treated as a universal ranking of the models. They do, however, show why agent speed cannot be reduced to output tokens per second. A capable model can still feel slow if it hesitates after reaching the answer, serializes work that could be parallel and repeatedly reloads context it already examined. The check-in experiment also suggests that harness design can materially change completion time without changing the underlying model.



