SWE-bench Pro was designed by the community after SWE-bench Verified became gameable — specifically, after it became clear that models trained on pre-cutoff data could score well by pattern-matching rather than by solving novel problems. The harder variant uses problems from post-training-cutoff repositories, which is why GLM-5.1's 58.4% score carries more weight than an equivalent performance on the original leaderboard would. The community that built the defense got beaten on the field it built the defense for. That closes the data-contamination objection that had been the first resort whenever an…