Dogfood Bench
Is Claude Code fixing its own issues faster?
Anthropic builds Claude Code with Claude Code. If AI makes software engineers more productive, the effect should show up in the claude-code issue tracker. This page measures it.
Caveat: this page is 100% vibe coded with Claude Code. If the numbers are wrong, that is also a data point.
Verdict, last 26 weeks
Loading data.
Fix rate
Issues fixed in a week, divided by issues opened in that week. The dashed line is the least-squares trend over the verdict window.
Closures by kind
Each week, issues opened (line) and issues closed (bars). Only "fixed" counts as a fix.
Open backlog
Open issues at the end of each week.
Fixed within 30 days
Of the issues opened in a week, the share fixed within 30 days. Weeks less than 30 days old are left out.
Median days to fix
For issues fixed in a week, the median time from open to close.
Method
A script reads every issue in the repository from the GitHub API. It records when each issue opened, when it closed, and the close reason. Each closure goes into one bucket by its reason. Who closed the issue does not matter, because the fixes supposedly come from AI anyway.
- Fixed
- Closed as completed, and not labeled invalid.
- Duplicate
- Closed as a duplicate of another issue.
- Not planned
- Closed as not planned, or labeled invalid. Most of these are bot sweeps of stale issues.
Only fixed issues count as evidence of work. The page uses complete UTC weeks only. Vertical lines mark model releases. The data refreshes daily. Last fetch: unknown.
Limits: "completed" is the default reason when someone closes an issue on GitHub. Some issues close as completed with no fix, for example when the reporter gives up. Issue volume also grows with the user count, so a falling fix rate can mean popularity, not decline. This is a half serious benchmark.