<!-- Merget benchmarks: merge conflict resolution results · https://merget.ai/benchmarks -->

# Fewer conflicts, less repeated work.

We benchmark the parts of our semantic merge queue that make it work: finding the right dependencies, resolving real merge conflicts and cutting repeated agent work. These are our findings from running Merget on open-source repositories.

Merget resolved all 223 conflicted files in two runs and found more reference calls than graphify in 11 of 12 repositories. On a Flask workload, its one-call resolver cut modeled spend by 69% against Merget with agent sessions alone, with the same branch outcomes.

Measured: Oct 13 – Nov 1, 2026

Code: Open-source repositories

Runs: Repeated on pinned builds

## Built to understand conflicts and resolve them efficiently. Each result links to its workload and measurements.

11/12repositories with higher call-graph recall than graphify

Higher recall in 11 of 12

### Call-graph accuracy vs graphify

More of the real dependency graph, with higher verified precision in 10 of 12 repositories too. We tested 12 open-source repositories across 9 languages.

Oct 13, 2026

[Open](https://merget.ai/benchmarks/graph-accuracy)

223/223conflicted files resolved in each of two runs

223 resolutions, repeated twice

### Conflict resolution with one model call

One model call per conflicted file, at about $0.018 modeled cost per file. All 214 checkable files kept Git’s already-merged text and contained no conflict markers, and about 84% matched the developers’ resolution.

Oct 30, 2026

[Open](https://merget.ai/benchmarks/conflict-resolution)

8.4×faster planning on Flask with 40 workers versus 1

40-merge plan in under 30 seconds

### Planning speed as workers are added

Merget planned 40 Flask merges in 28.9 seconds, down from 241.9 with one worker. On Ripgrep the speedup was about 13×.

Oct 29 – Nov 1, 2026

[Open](https://merget.ai/benchmarks/planning-scale)

69%lower modeled spend than Merget with agents alone

Same outcomes, fewer agent sessions

### Less agent work with the one-call resolver

On Flask batch 2, the resolver cut agent sessions from 7 to 2 and median time from 478.3 to 354.3 seconds, with the same branch outcomes. List-price spend was 57% lower.

Nov 1, 2026

[Open](https://merget.ai/benchmarks/merge-workloads#resolver)

## We ran these on real code. Open-source repositories, historical merges and repeated runs.

Real repositories

Every benchmark uses open-source repositories such as Flask, Ripgrep, Vue and TypeScript, at pinned commits.

Real merges

Conflicts and batches are replayed from each project’s own merge history, and resolutions are compared with what its developers committed.

Repeated runs

Timing, cost and resolution results come from repeated runs on a pinned Merget build, with the same model and budget on every side of a comparison.

## Give your agents a merge queue that helps finish the work. Start with evidence from your own repository.

These results are why we pair a detailed PR graph with a low-cost first resolution pass. More dependency context finds more collisions, and resolving files without a full agent loop cuts the work needed to finish a batch.

Start in Advisory mode to see Merget’s findings on your own pull requests.

Next step

Sign up free. Connect one repository and see where your parallel pull requests collide.

[Sign up free](https://sema.merget.ai/login)
