Gemini 4 Carbon, the model we covered earlier, has a new problem: in a private benchmark run by Kernel, it worked on bug fixes for nearly an hour while Claude Opus 5.5 and GPT-6 Astra finished the same kind of job in minutes. Kernel, a company that benchmarks model performance, published the analysis on its blog in November 2026. The results come from Kernel’s own evaluation, which it has not released, so treat them as one team’s measurement and not a verdict.

Each task hands the agent a codebase from just before a bug was fixed, plus a report describing only the symptoms. The agent has to diagnose the cause and patch it without touching anything unrelated. Kernel says that takes engineering judgment more than raw intelligence: picking the right code path, discarding rival explanations, and knowing when to stop.

Kernel first asked whether Carbon is simply slower per step. Part of it is. Provider queues, network lag, input handling, and output speed account for 16 to 30 percent of Carbon’s extra time against Astra, and 45 to 49 percent against Opus. The remainder comes from longer runs. Measured per bug, Carbon logged roughly 7x Astra’s step count and 11x its actions (a file read or a shell command, for instance). Its output ran 7x larger too, and its input volume topped 16x, since every step resends a longer conversation.

Kernel ruled out two easy explanations. Kernel also swapped in a pair of alternative coding agents, opencode plus mini-SWE-agent. Results barely moved, so the harness is not the culprit. Effort labels differ by provider, so Carbon’s “high” might equal a rival’s maximum. Kernel tried Carbon on low instead, and the model often delivered no fix, looping over files it had previously opened.

The trace labels show where the time went. Kernel tagged every step in the Astra and Carbon runs and found that about half of Carbon’s steps served goals outside the bug report. Once Carbon located the cause, it kept auditing neighboring code and hardened against cases the ticket had already excluded.

One sample trial makes the pattern concrete. Astra found the cause within a minute and had a patch drafted three minutes after that. Carbon found it within about six minutes, then wrote no code for another 25. Kernel wanted to know whether Carbon was unsure or just habitual, so it reran the trial with a prompt injected every five turns asking whether the model already had enough to make the fix. Carbon finished in 28 minutes instead of 37, and the wait from diagnosis to first edit fell from 25 minutes to about 10. Kernel reads that as a behavior problem, with a possible origin in training that rewards correct answers without penalizing extra steps. That origin is Kernel’s hypothesis, not something it demonstrated.

Even the necessary work ran slower. Astra averaged over two tool calls in each step and often opened several files together, while Carbon averaged about one. Carbon also repeated itself: 39 percent of its reading covered lines it had viewed before, against 7 percent for Astra. Kernel estimates those two mechanical habits explain roughly 25 percent of Carbon’s excess runtime.

The complaints were already public. On November 11, Theo of t3.gg posted a run of Carbon on low taking 22 minutes 21 seconds against 9 minutes 33 seconds for Opus 5.5 on high. Elliot Arledge, Shayan, and Simon Mahns posted similar frustrations between November 12 and 16, and the Kernel post drew 238 points and 125 comments on Hacker News. Kernel also priced the runs under several plans, using subsidization rates of 40 times for Opus, 13 for Astra, 3 for Carbon, and 17 for Hydrogen.

Kernel sells no competing model, but the benchmark is closed, and a task set built around minimal fixes rewards restraint by design. A team that picked Carbon for its price should time it on its own bug backlog, with a stopping prompt like Kernel’s, before committing a developer-facing workflow to it.

Kernel, “Why is Gemini 4 Carbon taking so long?”, kernel.dev blog, November 2026.