Anthropic Outlines Cost-Optimization Strategies and Commands for Claude Platform

Anthropic has provided guidance and updated the claude-api skill within Claude Code to help developers optimize cost and performance on the Claude Platform. The article, authored by Lance Martin, outlines three primary commands: prompt-audit, cost-optimize, and hillclimb.

The prompt-audit command scans workspaces for anti-patterns such as mandatory scratchpads or stale few-shot examples that degrade performance on newer models. In a customer support benchmark migrating from Opus 4.8 to Opus 5.5, the model switch alone reduced costs by ~18%, and removing anti-patterns via prompt-audit yielded an additional 9% cost reduction and a ~2 percentage point accuracy gain.

The cost-optimize command analyzes usage patterns and applies levers like prompt caching, output bounding, and batching. Benchmarks showed significant savings: LegalBench saw ~84% fewer thinking tokens and ~67% lower cost; tau2-bench retail saw ~73% lower cost with no change in pass rate; OfficeQA Pro saw ~72% lower cost with no significant score change; and SWE-bench Verified saw ~24% lower cost.

The hillclimb command performs an iterative search over model and effort choices, splitting evaluations into train and test sets to find optimal configurations. In a customer support benchmark, it reduced costs to one-fifth of the original setup while improving accuracy on held-out tickets.