PersonaCodeBench - SWE-bench Pro Adapted Tasks: leaderboard
Persona-adapted coding tasks requiring native software correctness and all mandatory evolved user requirements. Joint pass rate (%). PersonaCodeBench Lite100 fixed task–persona pairs, source-specific scaffolds under Harbor0.23.0; simulator/judge GPT-6 Luna medium. Joint pass requires both native completion and every mandatory personalization requirement. Evaluation errors/unexecuted tasks count zero. Per-row effort retained; GLM5.3 row scores textual part only. Runtime-specific outer watchdogs and source-specific inner limits remain as published.
Metric: Joint pass rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=max; source-specific scaffold) | 80.8 | |
| 2 | GPT-6 Astra (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=xhigh; source-specific scaffold) | 73.1 | |
| 3 | Claude Fable 5.1 (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=high; source-specific scaffold) | 69.2 | |
| 4 | Claude Sonnet 5.5 (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=high; source-specific scaffold) | 65.4 | |
| 5 | GPT-6.1 Sol (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=xhigh; source-specific scaffold) | 57.7 | |
| 6 | GPT-6 Astra (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=medium; source-specific scaffold) | 57.7 | |
| 7 | GPT-6.1 Sol (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=medium; source-specific scaffold) | 53.8 | |
| 8 | Claude Opus 5.5 (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=medium; source-specific scaffold) | 53.8 | |
| 9 | Claude Opus 5.5 (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=high; source-specific scaffold) | 53.8 | |
| 10 | GPT-6 Astra (Benchmark Sweep Run: source=PersonaCodeBench; Harbor0.23.0; effort=high; source-specific scaffold) | 53.8 |
Interactive version: theaggregate.ai/benchmark?slug=personacodebench-swe-bench-pro-adapted-tasks · How It Works · Data refreshed daily, snapshot 2026-11-06.