“DataCurve's testing on the DeepSWE benchmark identified a distinct failure pattern for Anthropic's Claude models, where they often missed stated requirements in multi-part prompts.”