MaxxBench tests coding models on familiar engineering tasks that ask for something different, questioning whether a high DeepSWE score reflects genuine capability. The benchmark aims to expose gaps between strong benchmark performance and a model’s ability to adapt when requirements change.