What AI benchmark scores actually mean.

Two kinds of score, what each can tell you, and how to read a gap between two models.

Facts below come from the datasets on this site, last updated 29 Oct 2026, 14:36 UTC. Interpretation is labelled as ours. We have not tested these products hands-on.

What the evidence shows

• Preference scores (Arena) come from people choosing between two anonymous answers. The top overall score among the listed models is 1,529 for Gemini 4 Argon, from 11,889 votes, with a 95% range of 1,522 to 1,535.
• Accuracy scores (Epoch) count right answers. The top SimpleQA Verified score among the listed models is 76.0% for Gemini 4 Argon, so even it did not answer about 24% of those questions correctly.
• Scores on different tests use different units and cannot be added together or compared across tests.
• Each Arena score has a range. If two models' ranges overlap, the data does not separate them with confidence.

Our interpretation

Treat a preference score as “people liked this answer style” and an accuracy score as “this many test questions were right”. Neither says the model will behave the same way on your task.

What to do next

See exactly what each test measures on How we test.