SWE-bench Verified: Claude Mythos Preview Leads at 93.9%, But SWE-bench Pro Reveals 70-Point Reality Gap

BenchLM.ai·medium signal

The SWE-bench Verified leaderboard now shows Claude Mythos Preview at 93.9%, followed by Claude Opus 4.7 Adaptive (87.6%) and GPT-5.3 Codex (85%), with 44 models evaluated. However, the more rigorous SWE-bench Pro tells a different story: top models score around 23%, compared to 70%+ on Verified, revealing a 70-point gap between benchmark performance and real-world capability on contamination-resistant evaluation. For builders evaluating coding agents, SWE-bench Verified scores are increasingly unreliable as a proxy for actual engineering performance.