Why is Gemini 4 Carbon taking so long?
TL;DR Highlight
Carbon took nearly an hour on bug fixes that Astra and Opus finished in minutes, mostly from extra unrequested work.
Who Should Read
A backend developer who plugged a new LLM into a coding agent and is puzzled by much longer run times and higher costs.
Core Mechanics
- Same bug fix took close to an hour
- On a private eval, Carbon spent nearly an hour on tasks Opus and Astra finished in minutes.
- The task: diagnose a bug from symptoms alone and fix it without unrelated changes.
- Inference speed explains only part
- Slower inference accounts for 16-30% of Carbon's extra time versus Astra.
- Versus Opus it accounts for 45-49%.
- The rest comes from longer trajectories.
- It simply does far more work
- Versus Astra, Carbon takes about 7x the steps and 11x the actions like reading files or running commands.
- Because the conversation is re-sent each step, it processes over 16x the input.
- Half the steps weren't in the ticket
- After finding the cause, it kept auditing code the fix never needed.
- It also hardened against cases the ticket had already ruled out.
- Check-ins cut 37 minutes to 28
- The authors injected "do you already have enough to make the fix?" every five turns in a modified harness.
- The gap from naming the cause to first edit fell from 25 minutes to about 10.
- They read this as a behavior issue, not a knowledge gap.
- Necessary work is inefficient too
- Astra made over two tool calls per step versus about one for Carbon.
- 39% of lines Carbon read were already seen, versus 7% for Astra.
- The authors estimate these two issues cause roughly 25% of the excess runtime.
- Not the harness or effort setting
- Running in opencode and mini-SWE-agent showed no meaningful difference.
- On low effort, Carbon often failed to produce a fix, re-reading code in circles.
Community Reactions
- Slow but good results, per users
- One commenter says it needs little supervision and is close to Opus 5.5 in quality.
- Another says it found optimizations other models missed in deep analysis.
- Extra work may be a feature, some argue
- One user uses Carbon for reviews because it flags bugs in distant code and finds dead code.
- Others counter that over-proactive behavior is annoying.
- Removing a harness prompt helped, one user found
- A commenter found their harness injected "proactively go beyond the user's request" each turn, and patching it out improved things.
- They suggest the habit may also be baked into training.
- Other benchmarks showed similar numbers
- In 146 battles, Carbon took 6m08s versus 2m04s and made 21 tool calls versus 11.
- It won 83 to 52 when voters waited, but 2 to 40 with early votes.
- A stopping rule worked around errors, one user said
- One commenter hit TURN_STEPS_EXCEEDED errors and fixed them with an effort budget or stopping rule.
- Another reported OpenRouter timeouts since the first token took over 5 minutes.
How to Apply
- Add a stopping rule to your agent prompt
- Today, add "once you find the cause, make only the needed fix and stop" to your coding agent's system prompt.
- Compare time and step count on the same task before and after.
- Inject a check-in every N turns
- Insert "do you have enough to make the fix?" as a user message every five turns in your agent loop.
- Log time from naming the cause to first edit to measure the effect.
- Log time, steps, and tokens in model evals
- When evaluating a new model, record time, steps, tool calls, and input tokens alongside accuracy.
- Compare total cost per task rather than list price.
- Audit your harness for proactive prompts
- Check whether your harness injects text like "go beyond the request" every turn, and remove it if so.
- Compare unrequested changes before and after.
Try It
# Paste into the system prompt or first instruction of Claude Code or a coding agent (simplified version of the article's check-in idea)
You are a bug-fixing agent. Follow these rules.
- Trace the report's symptoms to the root-cause code path.
- Once found, change only the files the fix needs.
- Do not add defensive code for cases the report's facts already rule out.
- Do not re-read files you have already read. Read multiple files at once.
- Every 5 tool calls, ask yourself "do I know enough to fix this now?" and if yes, fix it immediately.
Bug report: {{bug report}}
When done, list the changed files and one line on why each change was needed.
Next: compare time and tool-call counts for the same bug with and without these rules.Terminology
Related Papers
LLM assisted reverse engineering
Pointing an agent at binaries with skills and MD notes made years-long hacks feasible
Metrics-in-the-Loop: Complexity Feedback Reshapes Agent-Written Code
Injecting per-turn complexity reports into coding agents cut code erosion without changing correctness
How the Claude Code team thinks about verification
E2E, property and differential tests plus verification rules let Claude run in parallel.
Running coding agents without handing them my credentials
Credentials stay on the laptop; agents run commands there via Touch ID approval
Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use
Reverse engineering shows Windows 11 logs which coding agents you use and reports it as optional diagnostics
The Silent Cost of Conversation History in LLM Reasoning
Keeping full conversation history dropped accuracy from 0.6 to 0.1 within 20 turns; a verified state summary prevented it.
Anthropic upgrades Claude Code Projects to run parallel AI workers
Claude Code Projects now has a coordinator that splits work into parallel threads and opens PRs.
Show HN: AgentRun – open-source, Self-improving Managed Agents
Agents save notes and skills after each run, cutting code review cost to 18 cents