I agree. The “frontier level” open weights models benchmark really well but fall behind in real world performance