Every ranked AI model, best first.

What SCI is not. It is not a product review. It measures benchmark capability, not latency, price, context handling, or how a model feels in daily use. A single number also hides real specialization: a model can be frontier at code and mediocre at vision.

Preview models. Rows marked preview are vendor-designated preview or pre-release SKUs. Their scores come from published evaluations of that SKU and may not match the shipping product. A pre-release SKU whose released version is also ranked is folded into that row.

New models. Rows marked new were added to the registry automatically, by fixed intake rules, when a new model appeared in several independent sources. They are reviewed for at least 14 days before they count as approved (longer while a check still flags them), and are removed if the review finds them wrong. The mark sits on the configuration whose score the row shows.

Headline score. A model's row shows its highest-scoring configuration among those with 10 or more hard-benchmark results. A configuration with fewer is still listed when you open the row, but it does not set the headline: a score resting on a handful of benchmarks can swing by several places when one of them changes. A model with no configuration that well measured shows its best one.

Effort settings. A result measured at a stated reasoning effort counts toward that effort's configuration, whoever reported it. A lab's launch numbers at maximum effort sit beside independent runs at maximum effort, not with the model's default-effort results.

Evidence badges. lab-reported only means every result behind a row's headline score comes from the model's developer; thin evidence means that score rests on fewer than ten hard-benchmark results, because no configuration of the model has ten. Neither changes the ranking; such scores tend to move once independent evaluations arrive.

Uncertainty. Each daily refresh publishes point estimates; bootstrapped confidence intervals are computed on periodic manual runs. The evidence count on each row is the honest trust signal; treat thin rows as provisional.