Claude Opus 4.8 achieves 58% Pass@1 on the DeepSWE coding benchmark, trailing GPT-5.5 but leading on cost efficiency · Digg