Trace analysis identifies reasons for Gemini 4 Carbon's slower coding runs
A private software-engineering evaluation reported that Gemini 4 Carbon sometimes spent close to an hour on difficult bug-fixing tasks that Claude Opus 5.5 and GPT-6 Astra completed in minutes. Trace analysis attributed the gap to slower inference, longer trajectories, unnecessary auditing, repeated file reads and fewer tool calls per step. Carbon took about seven times as many steps as Astra and eleven times as many actions in the analyzed runs. One displayed trial took 22 minutes 21 seconds for Carbon at low effort, compared with 9 minutes 33 seconds for Opus 5.5 at high effort.
The analysis identifies specific inference, behavioral and mechanical factors associated with Carbon's high latency on difficult coding-agent tasks, including unnecessary work and repeated file reads.
Why is Gemini 4 Carbon taking so long?
November 2026
While benchmarking Gemini 4 Carbon on one of our private evals, we found something surprising: it spent close to an hour completing tasks that routinely took Opus and Astra only a few minutes:
Time and cost per trial
Left is faster; lower is cheaper.
API list price On a $200 plan, fully used
#Gemini 4 Carbon#Claude Opus 5.5#GPT-6 Astra#Gemini 4 Hydrogen#opencode#mini-SWE-agent