HarnessAudit-Bench - Runtime Robustness: leaderboard

Metric: Perturbation stability under tool or runtime errors and noise (0-1; GPT-5.4 rubric score, one variant on each of 105 tasks). Source: arxiv.org. Saturation forecast: Rough model projection: around 2027. 10 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.6 (Claude Code)0.47
2Claude Opus 4.6 (Claude Code)0.38

Interactive version: theaggregate.ai/benchmark?slug=harnessaudit-bench-runtime-robustness · How It Works · Data refreshed daily, snapshot 2026-09-26.