What happened
Gemini 4 Carbon needed close to an hour on bug-fixing tasks from a private evaluation that Opus and Astra typically finished in a few minutes, Kernel reported. In each task, an agent receives a code repository captured just prior to a bug fix, plus a report describing only the symptoms, and must diagnose and repair the problem without making unrelated edits.
Slower inference explained 16 to 30 percent of Carbon's extra time against Astra and 45 to 49 percent against Opus. The remainder came from longer runs. Measured against Astra, Carbon's runs involved about sevenfold the steps and elevenfold the actions, roughly seven times the written output, and more than sixteen times the input processed. Running it in opencode and mini-SWE-agent made no meaningful difference, and on a low effort setting it often produced no fix.
Labelling each step showed that about half of Carbon's went to work outside the ticket, such as auditing unrelated code and guarding against cases already excluded. When a modified harness prompted it every five turns to consider whether it could already make the fix, one trial fell from 37 to 28 minutes. Carbon also averaged about one tool call per step versus more than two for Astra, and 39 percent of the lines it read were repeats, against 7 percent for Astra.
Context
The analysis followed public complaints from users who said Carbon was too slow for everyday use, even as some argued it was underrated. Comparisons are complicated because reasoning effort settings are not standardized across providers. The benchmark tasks also lean heavily on engineering judgment, such as tracing symptoms to the right code path, ruling out rival explanations and knowing when to stop.
The briefing
The week in tech, in one email.
The stories worth your time, every week. Unsubscribe anytime.
