Frontier models take the digital SAT, scored exactly like a student would be.
⚠ Showing sample fixture data — real benchmark results haven't been published yet.
CI bar spans the 95% bootstrap confidence interval; the
gold tick marks the point score. Bars share one axis, so overlap is visible directly.
Bands: models whose confidence intervals overlap (transitively) share a band and are
statistically tied — ordering within a band is noise, not signal.
Shuffle Δ: score change when answer options are shuffled. Large drops suggest brittleness.
Footnote rows show per-model API errors and
unparseable answers — nothing is silently dropped.