llm-cost-optimizer
Analyze and reduce LLM API costs through model routing, caching, and prompt optimization. TRIGGER when: user asks about LLM costs, API spend reduction, token optimization, model routing, or prompt caching. DO NOT TRIGGER when: user asks about model quality comparison, fine-tuning, or general prompt engineering.
LLM Cost Optimizer
Reduce LLM API costs systematically without sacrificing output quality.
Three Modes
1. Cost Audit
Assess current spend and find the 80/20 opportunities.
- Instrument -- Add token counting and cost tracking per request. Log model, input tokens, output tokens, latency, and use case.
- Find 80/20 -- Identify which 20% of use cases drive 80% of cost. Sort by total spend, not per-request cost.
- Classify -- Tag each use case by complexity: simple (classification, extraction), medium (summarization, Q&A), complex (reasoning, code generation, multi-step).
2. Optimize Existing
Apply techniques to reduce cost on current workloads.
- Routing -- Route simple tasks to cheaper/smaller models. See optimization-techniques.md.
- Caching -- Cache repeated or similar queries. Prompt caching for system prompts.
- Compression -- Reduce prompt size without losing quality. Trim examples, remove redundancy.
3. Design Cost-Efficient
Build new systems with cost awareness from day one.
- Budget envelopes -- Set per-feature monthly cost budgets. Alert at 80%.
- Routing layer -- Default to cheapest model that meets quality bar. Escalate on failure.
- Observability -- Track cost per user, per feature, per model. Dashboard with trends.
Optimization Order
Apply techniques in this order (highest impact first):
- Model routing (60-80% reduction potential)
- Prompt caching (40-90% on cached portions)
- Output length control (20-40%)
- Prompt compression (15-30%)
- Semantic caching (30-60% hit rate)
- Request batching (10-25%)
See optimization-techniques.md for detailed guidance on each technique.
What You Get
- A cost audit identifying which use cases drive the majority of LLM spend, with per-request and total cost breakdowns.
- Prioritized optimization recommendations (model routing, caching, compression) ranked by reduction potential.
- Concrete implementation steps with expected cost savings percentages for each technique.
Rules
- Never optimize before measuring -- instrument first
- Never sacrifice quality silently -- A/B test every change
- Cost per request is misleading -- optimize total spend per outcome
- The cheapest token is the one you never send