# @gleech — 2026-02-26

♥27 ↻0 · https://x.com/gleech/status/2027080422614585537

@davidad Even a conservative estimate of Alibaba benchmaxxing would falsify that. (e.g. Qwen2.5 dropped ~15pp OOD.) I think it's true with s/Sonnet 4/Sonnet 3 though.

tags: author:gleech, kind:tweet, model:claude-3-sonnet, model:claude-sonnet-4, thread-context, year:2026