← All benchmarks

SciCode (Fugu background, patched libraries)

Science

Sakana Fugu Ultra's source-specific SciCode evaluation of the canonical 288 test subproblems using Inspect AI with scientist-provided background. The report follows the Artificial Analysis implementation but upgrades numpy, scipy and sympy to avoid package-version test failures; exact package versions and repeated-run settings are not specified. Kept diagnostic because equivalence to the official no-background protocol or a versioned Artificial Analysis harness has not been established.

Top models (higher is better)