The top models, field by field.

Each field is decided by one benchmark, named in that field's header, and every model listed in it reported that same benchmark. Models are ranked within the field and given a percentile — the share of the field they are at least as good as. A model's overall position is the average of its percentiles across the fields it reports. Percentiles rather than raw ranks, because a field with 83 reporting models and one with 15 would otherwise reward the same achievement very differently.

Because averaging a whole group would rank on which test a lab chose rather than on capability. HumanEval is saturated near 92% while SWE-bench Verified sits around 80% for the same class of model, so a model that published only the easy one would float to the top of Coding. One deciding benchmark per field means every model in a list sat the same test. The other benchmarks in each group are still shown on model pages and in side-by-side comparisons.

A model has to report at least 3 of the seven fields to be ranked. Below that, one strong benchmark would outrank a model measured across six, which would make the table misleading. Models under the threshold still have their own pages and still appear in the per-field lists below.

No. It is a measure of which models published the best numbers on the tests they chose to publish. That is a real signal and a limited one — a lab can decline to report a benchmark it does badly on, and no one here re-ran anything. Read the per-field lists before the overall table; they are closer to the truth.

Reasoning, Maths, Coding, Knowledge, Multimodal, Instruction following and Human preference. Each groups the benchmarks that measure the same thing, and a model's field score averages the ones it reports in that group.