Atlas

Benchmarks

← All benchmarks

RiemannBench (Anthropic corrected grading, with tools)

Math · 2026-07-24

Anthropic's July 2026 evaluation of the 25 RiemannBench problems with tools at maximum reasoning effort, using corrected reference answers and grading. Reported accuracy averages four attempts per problem. Equivalence with Surge's original unrestricted-agent grading has not been established; this measurement remains diagnostic.

Top models (higher is better)

ModelScore
Claude Opus 579.0
Loading Atlas data…