What Claude Code's multi-agent code review costs and finds
By Steven Van ·
Sonnet 5.5's max code review matched 23 of 27 serious issues in 56 minutes, while a single agent matched 21 in 40 seconds.
Quickchat has published a benchmark of Claude Code's local /code-review command, comparing its one-agent and multi-agent settings with Quickchat Review, the reviewer the company built. The test used 20 pull requests from Martian's Code Review Bench, and the Claude Code agents ran on Sonnet 5.5.
What was tested
The post tested Claude Code 2.1.303's local command, not Anthropic's managed Code Review service. It ran headless in a Docker sandbox with no GitHub token. The high setting starts one agent. The max setting fans out to about ten finder agents, a verifier agent per candidate, then a sweep. To keep max affordable, every subagent was forced onto Sonnet 5.5, each review was capped at $30 and given 90 minutes. The author notes that none of these is a shipped default.
The 20 PRs came from Sentry, Grafana, Cal.com, Discourse and Keycloak. They were the ones where Quickchat Review on two other models had done worst, a selection the author expects to favour every other reviewer. Scoring used Martian's pipeline with one LLM judge, GPT-6.1 Sol.
Results for one agent and the fan-out
- Serious issues matched: 23 of 27 for max, 21 for high. The author says a two-issue gap is about the scoring noise and cannot say whether it would repeat.
- All answer-key issues matched: 47 of 65 for both.
- Unmatched claims per PR: 23.9 for max, 11.9 for high.
- Time per finished review: 56 minutes for max, 40 seconds for high.
- Recorded cost per finished review: $20.95 for max (19 of 20 finished), $0.11 for high.
Cost did not track PR size. Finished max reviews cost $15.17 to $26.09 each. Five of the 19 launched no verifier, and four of those named the budget line in their context as the reason.
Behaviour on one small PR
On browser-use #857, which has five defects in the answer key, high on Opus 5.5 found all five in 70 seconds for 32 cents. max with Opus 5.5 on every agent had spent $40.64 by the author's recount when the 30-minute limit stopped it, and printed nothing because Claude Code prints the review only when it finishes. With Sonnet 5.5 forced onto every agent, max cost $13.31 and found all five. With Haiku 5.5 it cost $0.53 and also found all five. Each was a single run.
Comparison with Quickchat Review
Two AI judges compared the reviews head to head. On Opus 5.5, their ratings did not separate Quickchat Review from the fan-out. It cost $4.61 and took 8.2 minutes per pull request, which the post puts at under a quarter of the fan-out's price per finished review and about a seventh of its wall-clock time under the author's concurrency and rate limits.