Each benchmark -> 0-100 percentile (rank among models that ran it); score = the plain mean of a model's observed percentiles. No imputation, no mixed units. The score IS the number you read.
Showing 17 models that ran ≥3 of 7 benchmarks (9 excluded for thin coverage). Price = median effective $/1M tokens; Speed = throughput + latency. The n/7 chip = how many benchmarks back the score. English = English-language intelligence, a background prior anchoring every Score (not a German benchmark, never in the German tables).