RiemannBench (November 2026 rescoring)
Math
Surge's RiemannBench board as rescored between its November 1 and November 3, 2026 site builds: unconstrained research agents with coding tools and web search on 25 research-level math problems. Surge's launch post says it gave models unlimited time per problem instead of one hour and tested during periods of more stable performance, but dates neither that change nor this rescoring, and the page has no changelog. Every published value is a multiple of 0.8, which fits five attempts on each problem. Kept apart from the original board because scores moved by as much as 25.6 points (Muse Spark 1.3 xHigh, 28 to 53.6).
Benchmark source: Surgehq