Methodology

Models are measured on different benchmarks, and a score on one is not comparable to a score on another. Averaging them would produce a single confident number with no meaning, so the table instead picks one reference benchmark per tab — the one the most models in that tab have been measured on — names it in the column header, and ranks only the models that have a score on it. Everything else is listed below the table with the reason stated.

Coverage beats health when picking that benchmark. Ranking on a pristine benchmark two models share tells you less than ranking on a tired one that forty-nine share. Where the chosen benchmark is anything less than active, the header says so and names the healthier benchmark it beat, so you can disagree with the trade-off rather than be subjected to it.

Read a ranking here as “best on this benchmark”, never as “best”.

A vendor-reported number and an independent measurement are two different claims about the world, and this site never averages them into one. Where both exist, the independent number is shown and the row says both exist. Where only the vendor has reported, the row says that too. 10 scores currently carry it.

The same applies to a model’s configurations. A model run at high reasoning effort and the same model run at low effort are two measurements, kept apart in the data. The leaderboard shows one row per model by default, carrying its best configuration, and says which one and how many it beat — a collapse that names itself rather than a merge that hides.

An agentic score is a system score. The harness is half of what was measured, so two rows for one model under different scaffolds is correct, not a duplicate.

A name that does not resolve to the registry is never guessed at. It is written to a review queue with its source and excluded until a human matches it. A near-miss that resolved silently would credit the wrong model with someone else’s score, and nothing downstream would ever catch it.

A row whose source contradicts itself is dropped and recorded — Epoch ships rows naming one model in the machine-readable column and a different one in the display column, and there is no rule for deciding which half is right.

A measurement the source reports two different values for is not dropped. Both values are shown, both flagged disputed, because picking one would be a coin toss presented as a fact. That currently affects 416 scores.

13 of the 80 benchmarks catalogued here have any score at all, across 376 text models and 84 image models. The gap is a licensing constraint more than a technical one: the most complete commercial source for every modality does not permit redistribution, so it is excluded entirely rather than used quietly.

The benchmarks section catalogues all 80 with their health, including the ones nothing here has a score on.

There is no database and nothing is fetched while you read this page. A scheduled job runs daily at 06:00 UTC, validates each source against a schema, writes a dated snapshot into the repository and commits it. The commit is what deploys the site, so every number on it is reproducible from a specific commit, and an upstream change is visible as a diff rather than a silent edit.

A source that fails keeps its last good snapshot and goes visibly stale rather than blanking. A source returning less than half its previous record count is treated as broken rather than believed.

• benchwikifetched 2026-10-19
• epochfetched 2026-10-19
• epoch-ecifetched 2026-10-19
• arena-mirrorfetched 2026-10-19
• openrouterfetched 2026-10-19

It will not show a number without a source and a date. It will not present a routed price as a vendor’s list price. It will not render a missing value as a dash or a zero, because both read as data. It will not ingest a source whose licence does not permit redistribution, however much better that would make the coverage look.

Where those rules cost something — a thinner ranking, a visible gap, a benchmark with no scores — the cost is shown rather than hidden.

The small marks beside creator names are each vendor’s trademark. They are here to identify who made a model and imply no endorsement of this site by any of them, or of any of them by this site. A creator with no mark shown gets its initial instead; that says nothing about the model. A model released jointly by several organisations is never given one organisation’s mark.