Normalization
For each benchmark, the score is scaled to 100: (score − floor) / (best − floor). The floor is the random-guess level (25% for GPQA, 10% for MMMU-Pro, 1000 Elo for GDPval, 0 otherwise). The best tracked model gets 100.
EditionUpdated daily
About the ranking
How ERAZ points are calculated from published results, and where the data comes from.
How points are calculated
Categories
The 14 benchmarks are grouped into 7 categories. Within each, the average is weighted: the most discriminating tests (HLE, FrontierMath Tier 4, FrontierSWE, GDPval) weigh 1.5×, ARC-AGI-3 1.25×, saturated tests (GPQA) 0.5×, the others 1×.
Overall score
Weighted average of the categories where the model has results. Default profile: Reasoning, Math, Code, Agents ×1; Knowledge ×0.75; Long context and Multimodal ×0.5.
Coverage
A model that only publishes its best results should not win. Final points equal score × (0.75 + 0.25 × coverage) × 10, i.e. 1000 points at most.
Why not SWE-bench Verified, MMLU or AIME?
These tests are saturated (over 95% for the best models) or contaminated: they no longer separate frontier models. The Artificial Analysis index is shown for reference but excluded from the calculation, since it already aggregates several of the benchmarks used here.
Categories in detail
Each category score (Reasoning, Mathematics, Code…) is a weighted average of its benchmarks, scaled to 100. Here are, for each one, its benchmarks, their weight, their floor and the current best result.
A complete worked example
The actual calculation of a model's points, with today's data. Pick a model:
Sources
Every score on a model page also links to its original source ("src" link).