Benchmarks / PR Arena
PR Arena
Built by aavetis · released 26 May 2025
A public tracker that counts pull requests opened on GitHub by coding agents (Copilot, Codex, Cursor, Devin, Codegen, Jules) and how many of them were merged. It uses GitHub search queries on branch prefixes and bot authors and refreshes several times a day.
Frontier
96%
Merge rate of ready PRs
Copilot · agent: GitHub Copilot coding agent
23 Sep 2026 · Source: aavetis (benchmark maintainers)
README "Current Statistics" table: 1,548,258 merged of 1,612,469 ready PRs. Codex was at 89.65% (6,343,777 of 7,075,934), Jules 85.47%, Devin 61.94%, Codegen 61.6%. Cursor showed 101.87%, which is a counting artifact, so it is excluded.
No fixed reference point (for example Elo scores or field signals).
What it measures
The share of an agent's non-draft ("ready") pull requests on public GitHub that were merged. The maintainer uses ready PRs rather than all PRs so that agents that open drafts first (Copilot, Codegen) can be compared with agents that open ready PRs directly (Codex). Counts come from GitHub search, for example is:pr head:copilot/ and is:pr author:devin-ai-integration[bot].
Merge rate of ready PRs: Merged PRs divided by non-draft PRs, from GitHub search counts, across all public repositories. Higher is better.
No fixed reference point, so we do not rate its status. Field signal. Merge rates reflect who uses each agent and where, not a controlled task set.
Facts
- Grain
- Task level: bounded tasks with a clear spec
- Environment
- field-data
- Grading
- field-signal
- Reuse
- No license file in the repo. Data is derived from public GitHub search counts. (unknown)
Limits to keep in mind
- Counts depend on GitHub search heuristics (branch prefixes such as head:copilot/ and bot authors). Agent PRs that do not match these patterns are not counted. Source
- The merged count can exceed the ready count. On 2026-09-23 Cursor showed 910,955 merged against 894,225 ready PRs (a 101.87% rate). The maintainer does not explain the cause. We do not show Cursor as the frontier for this reason. Source
- Merge rate is a selection signal, not a quality score. Agents that require a human to mark a PR ready (Copilot) get a high ready-PR merge rate because humans filter first. Source
Recorded results
A selection that shows the frontier over time. The full leaderboard is at the source.
Timeline
- 26 May 2025 — PR Arena starts tracking agent pull requests on GitHub. Source
Where it sits in the atlas
Go to the source
- Website prarena.ai
- Full leaderboard prarena.ai
- Code github.com
- Dataset github.com
Last checked 23 Sep 2026 against 5 primary sources, with a second independent check. See an error? Tell us.
How to cite
Credit the original work first: PR Arena by aavetis.
Then, if you used this page:
Can Agents Work. "PR Arena: frontier results and sources." https://canagentswork.com/benchmarks/pr-arena/ (accessed 2026-09-24). CC BY 4.0.@misc{caw-pr-arena,
title = {{PR Arena: frontier results and sources}},
author = {{Can Agents Work}},
year = {2026},
howpublished = {\url{https://canagentswork.com/benchmarks/pr-arena/}},
note = {Accessed 2026-09-24. CC BY 4.0}
}