RiemannBench (Anthropic no-tools protocol)
Math · 2026-06-09
Anthropic's June 2026 evaluation of the 25 RiemannBench problems at maximum reasoning effort, without coding tools or web search. Reported accuracy averages four attempts per problem. This protocol is distinct from Surge's unrestricted-agent evaluation and remains diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 55.0 |
| Claude Mythos Preview | 43.0 |
| Opus 4.8 | 34.0 |