Gemini 4 Carbon took about seven times as many steps as GPT-6 Astra in Kernel’s test
In Kernel’s own debugging test, Google’s Gemini 4 Carbon spent close to an hour on tasks that Opus and GPT-6 Astra finished in minutes. Slower inference explained only part of the gap, Kernel said, and half of Carbon’s steps went into work the task did not require.
Kernel writes on its kernel.dev blog that it ran the comparison with its own unpublished test. In the test, a coding agent receives code in the state immediately before a bug was fixed, together with a report that describes only the bug’s symptoms. The agent must find the cause and fix it without making any other changes.
Kernel first separated out inference: how long a single step takes, including queuing and text generation. Slower inference accounted for 16–30 per cent of Carbon’s extra time compared with Astra, and 45–49 per cent compared with Opus. The rest came from Carbon taking more steps in its runs. On average, Carbon took about seven times as many steps as Astra and 11 times as many actions, such as reading files and running commands. At every step the model receives the entire, growing conversation again, so Carbon processed more than 16 times as much input.
By Kernel’s classification, half of Carbon’s steps went into work the report did not ask for. After finding the cause, Carbon extended its checks to every related path and to code the fix did not need to touch. It also guarded the code against situations the report had already ruled out. In addition, Carbon made about one tool call per step, while Astra made more than two. Of the lines Carbon read, 39 per cent had already been read, against 7 per cent for Astra. Kernel estimates that these two inefficiencies explain about a quarter of Carbon’s extra time.
In one example run, Carbon found the cause in about six minutes but did not start editing code until 25 minutes later. Kernel modified the harness, the program that steers the agent, gives the model its tools and runs its steps. Every five turns, the modified harness asked whether the model already had enough information for the fix. The run was cut from 37 minutes to 28. For the experiment, Kernel re-ran one of the earlier runs. In Kernel’s reading, the slowness appears to stem from how Carbon works, not from a lack of information.
Kernel ran Carbon at high reasoning effort and Astra at its highest level. Reasoning effort levels are not directly comparable across companies, Kernel said. At low effort, Carbon often left the bug unfixed. When Kernel switched the harness to opencode or mini-SWE-agent, the result did not change significantly.
Inference speed also varies. On Google’s developer forum, a paying user reported that 12 of the user’s 75 Carbon responses in the Gemini Developer API took more than two minutes on 17 November. During the slow period, Google’s status page showed all services as operational, the user said.
According to Google’s guide, Carbon is available to verified enterprise customers on Gemini Enterprise Agent Platform in select regions.
Read more: Models