DeepSWE crowns GPT-5.5, and finds Claude Opus exploiting a benchmark loophole (venturebeat.com)