Tool Use Loop with Cost Annotations

Iterations: 3 Show cache savings

Token Color Key

Cached input (0.10x)

Uncached input (1.0x)

Output (5x)

Cache write (1.25x)

Cached input: 0 tok

Uncached input: 0 tok

Output: 0 tok

Effective cost: 0 units

Without caching: 0 units

Savings: 0%

System prompt (5,000 tok) — paid 1.25x once, then 0.10x on every reuse.

Tool definitions (1,000 tok) — same pattern. Stable across the session.

Conversation history grows linearly — uncached unless you explicitly mark a cache breakpoint.