Of 604 violations, Peer caught 508
Three judges found 604 violations in 276 open-source TypeScript files, at least two agreeing and none disagreeing. Each square is one; a lit square is one the reviewer caught.
Claude Haiku, Sonnet and Opus 5.5, each given the file and a short review prompt, caught 134 to 200.
And more of it real
Catching more did not cost precision: 83% of Peer’s findings were real, against 44% to 55% for Claude.
- Peer
- 107
- Haiku 5.5
- 257
- Opus 5.5
- 180
- Sonnet 5.5
- 111
Where each reviewer sits
Accuracy against cost and against time. The shaded corner beats every Claude model on both at once; only Peer is in it.
F1 balances what was caught against what was real. Cost at list price as a bare API call for Claude; Peer at its price of $0.10 per 1M tokens.
Tokens, money and time
What one review of one file takes, reviewer by reviewer.
- Peer
- 501
- Haiku 5.5
- 2,174
- Opus 5.5
- 1,598
- Sonnet 5.5
- 1,423
- Peer
- 19,960
- Haiku 5.5
- 1,558
- Opus 5.5
- 69
- Sonnet 5.5
- 181
- Peer
- 0.54 s
- Haiku 5.5
- 5.97 s
- Opus 5.5
- 5.92 s
- Sonnet 5.5
- 4.58 s
Not just style
Some violations come down to style conventions. Set those aside and Peer still catches the most, with the fewest false alarms.
- Peer
- 76%
- Haiku 5.5
- 40%
- Opus 5.5
- 31%
- Sonnet 5.5
- 26%
- Peer
- 0.20
- Haiku 5.5
- 0.74
- Opus 5.5
- 0.54
- Sonnet 5.5
- 0.37
Every number
Each accuracy figure with its 95% interval, from a bootstrap over files.
| Reviewer | Caught | Precision | False alarms a file | Tokens read | Written | $ per 1,000 files | Seconds, median | Slowest 5% |
|---|---|---|---|---|---|---|---|---|
| Peer | 84% 81%–87% | 83% 80%–86% | 0.39 0.32–0.46 | 501 | 0 | $0.05 | 0.54 | 0.66 |
| Claude Haiku 5.5 | 33% 30%–37% | 44% 39%–49% | 0.93 0.82–1.05 | 1,112 | 1,062 | $0.64 | 5.97 | 9.14 |
| Claude Opus 5.5 | 27% 23%–31% | 48% 42%–53% | 0.65 0.56–0.74 | 1,087 | 511 | $14.56 | 5.92 | 10.43 |
| Claude Sonnet 5.5 | 22% 18%–26% | 55% 48%–61% | 0.40 0.33–0.48 | 1,090 | 333 | $5.51 | 4.58 | 9.24 |
Run through Claude Code’s CLI instead of the bare API, the Claude reviews cost $1.44 to $46.42 per 1,000 files.
How we measured
- Files. 276 windows of code from ten of the most-starred open-source TypeScript repositories. Peer was never trained on them.
- Judges. Claude Opus, Claude Sonnet and GPT-6.1 Sol read every file apart. A violation counts where at least two agree and none disagrees. Two of them share a model with two reviewers, which favours Claude.
- Peer read each file once through its hosted API, as your agent would call it. Claude got the file with line numbers and the prompt below, at medium effort, with no tools. Never a list of what Peer checks: your agent does not have one either.
- Grading. Each Claude finding was matched to the kind of issue the judges graded. A match on a confirmed violation is a catch; a finding on code the judges found clean is a false alarm.
You review one file a coding agent just wrote, before it is committed. Look for what should change: unnecessary or dead code, duplicated logic, needless complexity, unclear names, loose types, missing or wrong error handling, bugs, and style the file itself breaks. The file may be an excerpt of a longer one; do not flag what lies outside it. Report only what you would ask the author to fix.
Put Peer in front of your agent
Peer opens by invitation. Join the waitlist, then add it with the install line.