Cost per Task and the Levers That Move It

A typical task for your coding agent takes 40 model steps. Each step sends about 40,000 input tokens (of which a 12,000-token prefix — system prompt, tool descriptions, project instructions — is identical on every step) and produces about 300 output tokens. Assume illustrative prices of $3 per 1M input tokens, $15 per 1M output tokens, and cached input billed at 10% of the input price.

1. Compute the cost of one task without caching, then with the prefix cached on every step after the first.
2. Engineering wants tasks under $1.50 on average. Propose levers in priority order and estimate their effect.
3. Which metric would you actually optimise and report, and why is "cost per task" alone misleading?