ScarfBench - Focused Apps (Skills Prompt): leaderboard
ScarfBench (2026) tests coding agents on migrating 29 focused enterprise Java applications between Spring, Quarkus and Jakarta EE (174 directed tasks), each agent in its own harness with a skills-directory prompt. Score is the percent of behavioral oracle tests that pass after the migrated code is rebuilt and deployed, with failed builds or deployments scoring zero; higher is better.
Metric: Behavioral test pass rate (%): mean share of the application's behavioral oracle assertions that pass on the migrated code, 0 when the target build or container deployment fails, over the 174 directed Spring, Quarkus and Jakarta EE migrations of the 29 focused applications; one pass@1 run per task at temperature 0 with the per-task skills directory prompt, each coding agent in its own harness; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 15 | #99 |
| 2 | GPT-5.4 | 14 | #135 |
| 3 | GLM-5.1 FP8 | 14 | #233 |
| 4 | Qwen CLI + Qwen3.5-397B-A17B-FP8 | 11 | |
| 5 | Claude Opus 4.6 | 7 | #113 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=scarfbench-focused-apps-skills-prompt · How It Works · Data refreshed daily, snapshot 2026-11-21.