Atlas

Benchmarks

← All benchmarks

GPQA Diamond (Epoch average of eight attempts)

General QA

Epoch's GPQA Diamond mean accuracy over eight attempts per each of the 198 questions. The explicitly documented Grok 4 July 2025 run includes API failures as incorrect attempts; this repeated-run mean does not have a single-attempt binomial denominator.

Top models (higher is better)

ModelScore
Grok 487.0
Loading Atlas data…