A $0.25 Model Beat a $15 Model. The Secret Was the Skill, Not the Brain.

Tessl ran 880 evaluations across 9 models with and without agent skills. The result inverts what most teams assume about AI costs.

Haiku 4.5, Anthropic's cheapest model at roughly $0.25 per million tokens, scored 84.3% when given a well-crafted agent skill. Opus 4.7, the most expensive at roughly $15 per million tokens, scored 80.5% without one. Read that again. The cheap model with good instructions beat the expensive model without them. Haiku's 23.1-point lift from the skill was the largest of any model tested.

Here's what really caught my attention. When you load skills on the frontier models, they all converge: Opus 4.7 hits 94.5%, Opus 4.6 reaches 93.8%, Sonnet 4.6 lands at 93.3%. That's a 1.2-point spread. You're paying 60x more for Opus over Sonnet to get 1.2 points. The skill is doing the heavy lifting, not the model.

This lines up with what I'm seeing from vercel-labs/skills, which hit 15.2K stars today. Their CLI installs reusable instruction sets across 45+ AI coding agents with one command: npx skills add [repo]. It's becoming the npm for agent skills, a shared specification that works across Claude Code, Codex, Cursor, OpenCode. 92 contributors, 25 releases.

The practical implication for anyone running agents in production: stop optimizing your model tier first. Write better skills. Invest engineering time in the instructions, constraints, and domain knowledge you feed the model. Then pick the cheapest model that clears your quality bar. For most tasks, that's probably Sonnet or even Haiku, not Opus. The cost difference at scale is enormous.

I've been running my own pipeline on a mix of models for months, and this matches my experience. The quality of the prompt architecture matters more than the model behind it. The skill is the moat.