SWE-Bench Verified Hits 80.9% Ceiling: Top 5 Models Converge Within 1% While SWE-Bench Pro Shows 56.8% Reality Check

BenchLM.ai·medium signal

The SWE-Bench Verified leaderboard shows unprecedented convergence at the top: Claude Opus 4.5 (80.9%), Claude Opus 4.6 (80.8%), Gemini 3.1 Pro (80.6%), MiniMax M2.5 (80.2%), and GPT-5.2 (80.0%) are all within 1% of each other, up from ~65% a year ago. Meanwhile, SWE-Bench Pro tells a different story: GPT-5.3-Codex leads at only 56.8%, revealing that the 80%+ scores on Verified may reflect benchmark saturation rather than true agent capability. The 24-point gap between Verified and Pro scores suggests the industry needs harder, more realistic benchmarks to differentiate coding agents.