llm cost optimization audit

The problem

LLM bills often grow into unmanaged line items that eat product margins. Developers frequently mistake setting a configuration flag for actually saving money, leading to "optimized" prompts that are silently more expensive than the baseline.

What it does

• Profiles token usage across input, output, and cache creation to find the real cost drivers.
• Identifies silent cache invalidators like per-request timestamps or unsorted JSON that break prompt caching.
• Calculates the break-even point for caching based on reuse frequency and TTL windows.
• Audits model routing to find high-volume, low-complexity tasks running on over-powered tiers.
• Sizes estimated savings for request batching and context compaction before you write any code.

Frameworks & tools

Works with any LLM provider usage logs. Specific logic for Anthropic Claude API prompt caching and Batch API. Uses current pricing snapshots for ROI calculations.

Why this beats prompting it yourself

Generic prompts often guess at savings. This skill uses a "profile first" discipline to ensure recommendations are based on your actual token distribution. It accounts for hidden costs like the prompt-cache write premium, which can actually increase your bill if reuse is infrequent.

Use cases

• Diagnosing why prompt caching is resulting in zero read tokens despite being enabled.
• Evaluating the margin impact of moving high-latency background jobs to a Batches API.
• Reducing context waste in long-running agent loops or chat histories.
• Right-sizing models for specific tasks like classification or extraction without losing quality.

Known limitations

Does not perform quality benchmarks or "capability bake-offs" between models. Focuses on inference costs only, excluding database, hosting, or egress fees.