BenchLM stops zero-filling missing benchmark scores, and the methodology change matters more than the leaderboard.

Their continuously-updated registry, last refreshed July 14, now tracks 296 benchmark definitions across 10 categories (BenchLM). BenchAlign v5 now estimates capability from available evidence instead of converting missing benchmarks to zeros, and labels every result "Supported" or "Estimated." Which means every cross-model comparison published before this change systematically penalized models with sparse eval coverage. Current listing: Claude Fable 5 at 83, GPT-5.6 Sol at 81.5 on Supported evidence, MiniMax M3 leading open-weight at 68.8. The top entry, Claude Mythos 5 at 85.9, carries only Estimated confidence and shouldn't be treated as confirmed.