12 tactics to cut your LLM bill in half

A 50% cost reduction rarely comes from one heroic change. It's twelve 5% wins compounding.

The list#

• Downgrade to the smallest model that passes your evals
• Prompt cache the system message
• Shorten the system prompt (measure eval delta)
• Cap max_tokens aggressively
• Use response_format: json_schema to eliminate retries
• Deduplicate identical requests with a cache
• Batch embeddings 100 at a time
• Use TokenBaazar's cheaper model routing for non-critical paths
• Stream to abort early on user cancel
• Compress RAG context (extractive summarization pass)
• Move cold prompts to Gemini Flash
• Audit your top-10 highest-cost endpoints monthly