Tip: Agent Skills Matter More Than Model Choice — 880-Eval Benchmark Shows Haiku 4.5 + Skill Beats Baseline Opus 4.7

Tessl·high signal

Tessl ran 880 evaluations across 9 models (Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5, three Codex variants, Cursor Composer-2) with and without agent skills. Result: Haiku 4.5 with a skill scored 84.3%, beating baseline Opus 4.7 at 80.5%. Haiku's 23.1-point lift was the largest of any model. With skills loaded, the frontier models converge: Opus 4.7 (94.5%), Opus 4.6 (93.8%), Sonnet 4.6 (93.3%) — within 1.2 points. Takeaway: invest in skill engineering before upgrading models.