AI NewsProductivityReported
Gemini 4 Carbon takes seven times the steps of Astra, Kernel finds
Kernel measured Carbon taking seven times as many steps as Astra on the same bug-fixing tasks.

Why it mattersCarbon costs less per token and still costs more per task. A buyer picking a coding model by listed price will pay more. The step count is the number to watch.
A coding model that costs less per token can still cost more per task. Kernel, a research group writing on its own blog, ran three models on the same bug-fixing tasks. The models were Google's Gemini 4 Carbon, OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5. Carbon took about an hour on jobs the other two finished in minutes. Most of the gap, Kernel says, comes from Carbon doing more work per task.
The tasks came from Kernel's private evals. Kernel also ran Theo Browne's public benchmark. Gemini 4 Carbon low finished in 22 minutes 21 seconds. Claude Opus 5.5 high finished in 9 minutes 33 seconds. Each agent got a codebase from just before a bug was fixed. The report named only the bug's symptoms. The agent had to find the bug and fix it, without changing other code.
Three causes, measured separately
Kernel pulled the extra time apart into three parts. The first is slower inference. Carbon waits in its queue and takes longer per step. That accounts for 16 to 30 percent of the gap against Astra, and 45 to 49 percent against Opus. The second is unnecessary work, the biggest share. Carbon checks unrelated parts of the code. It adds guards against scenarios it has already ruled out. Kernel marks about half of Carbon's steps as unasked work. The third is mechanical inefficiencies, about a quarter of the extra runtime. Carbon makes fewer than one tool call per step. Astra makes more than two. Carbon also re-reads the same lines of code on later steps.
The scale of the extra work is large. Carbon takes seven times as many steps as Astra on the same task. It makes eleven times as many tool calls. It writes seven times as much output. It processes sixteen times as many input tokens. The re-reading is measurable. Of the lines Carbon reads back, 39 percent are duplicates of lines it has already read. For Astra the figure is 7 percent.
The check-in experiment
Kernel wanted to know why Carbon kept working. Was it short on information, or did it keep looking after it had enough? The team changed its harness to ask Carbon every five turns: do you have enough to make the fix? With that one change, Carbon finished one trial in 28 minutes instead of 37. The wait between Carbon naming the cause and making its first edit fell from 25 minutes to 10. Kernel reads that as a behaviour, not a knowledge gap.
What a buyer sees
Kernel notes Carbon is now priced at its post-introductory rate. The step count and the token count decide the bill, not the price per token. A cheaper token can still cost more per task. The model writes, reads and calls tools several times more. The methodology is Kernel's own. The traces come from Kernel's harness, and the private evals are not published.
A team picking a coding model this quarter has one practical move. Measure the step count and the total token count on your own tasks. Do not pick by the published benchmark pass rate alone. The number that matters is how much work the agent chooses to do.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
