Pattern: The Skill > Model Inversion — Right Skill Flips the Cost-Performance Leaderboard

Tessl·high signal

Tessl's 880-eval study crystallizes an emerging pattern: with agent skills loaded, the cheapest model (Haiku 4.5 at ~$0.25/M tokens) outperforms the most expensive model without skills (Opus 4.7 at ~$15/M tokens). At the frontier with skills, all three Claude models converge within 1.2 points. This inverts the traditional 'pay more for better results' assumption. For builders optimizing cost: invest engineering time in skill quality first, then pick the cheapest model that clears your quality bar. The skill is the moat, not the model tier.