Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of Astra
A.I. / news
Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of Astra
Artificial Analysis puts Carbon at 36.0 minutes per Intelligence Index task against a 12.4-minute median, and a Kernel study traces most of the gap to extra steps.

Gemini 4 Carbon, Google's top-tier model, took 36.0 minutes per task at its High setting on Artificial Analysis's Intelligence Index, against a median of 12.4 minutes for comparable models, according to the benchmarker's model page. A separate study published by Kernel on Nov. 18 traced most of the gap to the number of steps the model takes, with generation speed accounting for a smaller share.
The Kernel post, which names no author, ran Carbon on a private bug-fixing eval. Each agent received a codebase from just before a bug was fixed and a report describing only the symptom. Carbon spent close to an hour on tasks that Claude Opus 5.5 and GPT-6.2 Astra usually finished in a few minutes, the post said.
What Artificial Analysis measured
Artificial Analysis lists Carbon's release as Oct. 28, 2026, with a 1M-token context window and API pricing of $4.00 per million input tokens and $20.00 per million output tokens. The page puts the median comparable model at $2.00 and $10.00.
Time to first answer token depends heavily on the effort setting. Output speed barely moves.
- Carbon High36 min
- Carbon Low30.1 min
- Carbon Medium29.6 min
- Comparable-model median12.4 min
Source: Artificial Analysis, Gemini 4 Carbon model page, accessed 2026-11-18
At High, the page lists 75.03 seconds to the first answer token, 57 tokens per second of output and an Intelligence Index score of 60, against a median of 43 for comparable models. Carbon ranks 7th of 92 on intelligence and 87th of 91 on performance, the page says. Google's other models on the same index score 54 for Gemini 4 Hydrogen and 50 for Gemini 4 Argon.
Where the time goes, per Kernel
Kernel attributed the gap to three things. Slower inference explains 16 to 30 percent of Carbon's extra time against Astra and 45 to 49 percent against Opus. The rest comes from longer trajectories.

Compared with Astra, per bug on average, Kernel measured the following for Carbon.
| Measure | Carbon vs. Astra |
|---|---|
| Steps taken | about 7x |
| Actions (file reads, commands) | about 11x |
| Input processed | more than 16x |
| Lines read that had already been seen | 39% vs. 7% |
In one sample trial, Astra identified the cause within a minute and wrote its fix three minutes later. Carbon identified the cause in about six minutes and wrote no code for another 25 minutes. Kernel said about half of Carbon's steps went to work the ticket did not ask for, such as auditing code beyond the bug.
Kernel also injected check-ins every five turns asking whether the model had enough to make the fix. Carbon then finished in 28 minutes instead of 37, and the wait between diagnosis and first edit fell from 25 minutes to about 10. The authors wrote that "the slowness seems to come from how Carbon behaves, not what it knows."
Astra made more than two tool calls per step, often reading several files at once. Carbon made about one. Kernel puts those habits at roughly 25 percent of the excess runtime.
What Kernel ruled out and what it did not test
Running Carbon in two other agent harnesses, opencode and mini-SWE-agent, showed no meaningful difference, Kernel said. On low effort, Carbon often failed to produce a fix and looped, re-reading code it had already read.
The post cites public complaints on X, including one from Theo (t3.gg) on Nov. 11 listing Carbon at its low setting at 22 minutes 21 seconds against 9 minutes 33 seconds for Opus 5.5 at high. Those are single runs quoted by Kernel. The eval itself is private, so the 7x and 11x ratios cannot be reproduced by outside parties.
Kernel's cost charts assume subscription subsidy rates of 40x for Opus, 13x for Astra, 3x for Carbon and 17x for Hydrogen, and the post lists Carbon at its post-introductory price.
Sources disagree on price: third-party aggregators list an introductory $2 and $10 rate. Google's own pricing page was not reviewed.
Neither Kernel nor Artificial Analysis says whether Google plans to change Carbon's default effort or its step behaviour.
For context on how Google's cheaper model sits in Search, see the earlier report on Gemini 4 Hydrogen. For the rival Kernel used as its baseline, see GPT-6.2 Astra and Sol pricing.
Sources
More in A.I.
- 01Researcher Says Claude Fable 5.2 Learned Protein Sequences; Test Shows a 45-Point GapXinming Tu's probe of 668 obscure bacterial proteins finds the newest Claude models predict real sequences far better than shuffled ones. GPT, Gemini and open-weight models do not.
- 02Tencent's Hy4-flash Weights Carry a License That Requires a Separate Deal Above $50 Million in RevenueThe 12B-active model is ungated on Hugging Face, but it ships under the Tencent Hy License, not Apache 2.0.
- 03OpenAI's GPT-6.2 Astra Lists at $8 per Million Input Tokens, 20% Under Its Predecessor, While Coding Scores Stay FlatArtificial Analysis ranks Astra second at 62 behind Claude Fable 5.2, and Sol's coding-agent score fell 3 points.
- 04Google Put Gemini 4 Hydrogen Behind Search for Everyone on Nov. 12, and Published No Click FiguresThe one randomised study of AI Overviews found a 39.8% drop in outbound clicks on early-2026 Search.