Score coverage explained: why no AI model in our League has a complete benchmark record.

By AI Pro League · Published Nov 4, 2026

Score coverage, in the AI model rankings on this site, is computed the same way for every model: add up the weights of the benchmark components the model has a recorded result for. The five components carry weights of 40% (LiveBench), 20% (SWE-bench Verified), 15% (GPQA Diamond), 15% (LMArena Elo) and 10% (MMLU-Pro), so a model measured on all five would show 100% coverage. None does. Across the 45 scored models in our model database, the deepest record is 90% — shared at the top of the table below — the median is just 40%, and a further 14 tracked models have no recorded component at all and carry no League Score.

What score coverage means — and how the score is built

A League Score is a weighted average of a model’s normalised component scores, divided by the weights of the components it actually has — never by the full 100%. A missing benchmark is skipped, not scored as zero: the score describes the record that exists, and coverage tells you how much record that is. That is why two models can sit side by side in the table with scores built from very different amounts of evidence, and why we print the coverage figure wherever the score appears. Each component is explained on its own page: LiveBench, SWE-bench Verified, GPQA Diamond, LMArena Elo and MMLU-Pro.

How many AI models have a complete benchmark record?

None — in our database, and the shortfall is structural, not a bookkeeping lag. The components are measured by different organisations on different schedules: LiveBench (run by the LiveBench project) has recorded a result for 31 of our 45 scored models, while MMLU-Pro results exist for only 8 and LMArena ratings for 10. A model can only reach 100% coverage by being measured on all five, and no lab has put its flagship through every one of them on the record we track.

The deepest records in the League

Depth and height are different things: the deepest records include both a top-five score and one of the table’s lowest, because being measured everywhere means having nowhere to hide. A model measured only where it is strong keeps a flattering average; a model measured everywhere shows its weak components too.

Worked example: the leader’s 55% vs a 90% record

The overall leader, GPT-6 Astra, scores 99.8 — computed from just two components (LiveBench overall, GPQA Diamond), a 55% coverage record: its weighted component scores are renormalised over 55 points of weight, not 100. Claude Opus 5 carries the League’s joint-deepest record at 90% (LiveBench overall, SWE-bench Verified, GPQA Diamond, LMArena Elo). Neither number is wrong; they are answers to slightly different questions. The 55% score says “on the two hardest-to-fake measures we track, nothing recorded beats it.” The 90% record says “this score has survived four different kinds of test.” Choosing between them is a judgement about evidence, which is exactly why the coverage figure is printed instead of buried.

When a League Score is one benchmark wearing a sash

22 of the 45 scored models — nearly half the table — rest on a single component. For those models the League Score is, arithmetically, that one benchmark’s normalised score: the renormalisation divides by the single weight present, so the headline number and the component number are the same. The top five single-component scores:

LiveBench-only scores dominate that list because LiveBench is the most-recorded component — 31 of 45 scored models — and at 40% weight it is also the floor for a single-component record. The thinnest records in the table sit at 15% coverage: one MMLU-Pro or LMArena result, renormalised into a full-looking score.

Does low coverage mean a bad model?

No — it means a young or selectively measured one. Coverage grows as independent benchmarks publish results: several of the 14 unscored models in our database are recent releases whose records are still empty, not failed. Read the pair of numbers the way the table prints them: the score is the performance on the evidence available, the coverage is how much evidence that is. A 90-point score at 15% coverage is a promising rumour; the same score at 90% is a finding. Our open-weight ranking and context-window analysis both lean on the same discipline: no number from this database is quoted without the record behind it.