shared older model numbers here https://www.latent.space/p/ainews-frontiercode-benchmarking

tldr theres been broad progress despite your observed regressions