Multimode AI · Benchmarks

Benchmarks

Honest tests of AI on hard, checkable problems. Every number here has been stress-tested against its own method, because the confident number is the thing to distrust. Measurement is easy. Measurement integrity is the whole job.

girih-bench preview

girih-bench

Fable 5 vs GPT-5.6 Sol

Can a frontier model rebuild a 1000-year-old Islamic quasicrystal? I scored it by physics, the way quasicrystals were actually discovered.

A tie on the physics, and a lesson in measurement integrity.

See the benchmark →