LLM Benchmarks

How the Index Is Built

Each benchmark is read as a placement rather than a raw score. Within one benchmark the best result sits at 100, the worst at 0, and every other model falls somewhere in between. The index is the weighted average of those placements over the benchmarks a model has actually run: a benchmark it was never run on is not counted as a zero, it is simply absent. Benchmarks do not count equally, and the weight each one carries is listed with it. A model needs results on at least two index benchmarks before it is ranked, so a single result places a model on one benchmark's line without giving it a standing of its own. Placements are used instead of raw accuracy because benchmarks differ sharply in difficulty and scale, and scores from different benchmarks cannot share an average. Weights: AIM-FinanceReasoning (10%) + AIM-Text-to-SQL (10%) + AIM-Agentic-RAG (10%) + SciCode (10%) + FrontierMath (10%) + Terminal-Bench 2.1 (10%) + CritPt (10%) + IFBench (10%) + MMMU-Pro (10%) + ARC-AGI-3 (10%).