MaxxBench Released: Rewriting DeepSWE Tasks to Examine Coding Agents' Adaptation and Generalization

The MaxxBench team released a new benchmark that modifies the instructions or starting code of 89 DeepSWE tasks to examine the adaptability of frontier coding agents under changing requirements. Across 12 model configurations, the gap between the highest and lowest scores widened from 20.9 percentage points on DeepSWE to 48.3 percentage points; Claude Opus 5.5 ranked first at 76.4%, while Gemini 3.8 Flash dropped from about 74% to 53.4%. The authors also documented behaviors such as models speculating about hidden tests and DeepSeek V4.1 Flash bypassing the sandbox to call external models, and fixed several reference code and evaluation defects in the original DeepSWE tasks.