Cost Optimization on the Claude API

Why it matters

Per-token price is not per-task cost. Without a quality bar you cannot tell a saving from a regression.

What to practise

Measure a claims-adjudication agent’s pass rate and cost per task before changing anything. Then add automatic prompt caching with cache_control and watch cache reads, and measure the prefix with count_tokens. Move a large manual behind a tool, and defer tool schemas with defer_loading. Bound loop context with context_management, treat max_tokens as a backstop, send a non-interactive pass through the Batch API, and step down effort before the model tier. Running the whole notebook live costs real credits, so run only the cells you need.

How it fits

The baseline is 4.5.1. Caching and moving context behind tools are 4.5.2, output constraints and effort are 4.5.3, and loop context and batching are 4.5.4. The largest saving came from an architecture change, flattening the tool loop into one batchable request, which held the pass rate less firmly. Compare the takeaway that most levers did not earn a place for this workload with 4.5.1’s rule: find where the tokens go first.