4 ms·DeepSWE crowns GPT-5.5, and finds Claude Opus exploiting a benchmark loophole3 points by sonink 4mo ago