Consensus models
262
58 disputed across boards
Empirical studies
Cross-benchmark findings computed directly from canonical views at build time. Each study states its exact method, limits, and excluded rows. Data build 17 Nov 2026.
262
58 disputed across boards
4.4
42% variance on first axis
106 / 131
Within 95% confidence overlap
91d
28 multi-group tracks
Latent Bradley-Terry capability scores, where human Arena preference diverges, domain fingerprints of generalists and specialists, and which model teams fix the most bugs.
262 models scored · 104 in Arena gap audit
Read study and methodPairwise rank correlation across boards, a clustered map of which benchmarks measure the same thing, and near-duplicate pairs with Spearman rho at least 0.90.
4.4 effective dimensions · 32 near-redundant pairs
Read study and methodMarginal accuracy per dollar and reasoning effort curves from minimal to maximum budgets, flagging non-monotone regressions.
332 effort curves (83 non-monotone)
Read study and methodFrontier progression over release dates, which lab held the lead and for how long, how fast the price of a fixed score falls, and the open-weight lag.
91 days median open-weight lag
Read study and methodPer-model evidence grades (A to D), the probability that each model holds each rank, single-run flags, and vendor self-preference audits.
71 Grade A models · 106 boards tied at #1
Read study and method