AI Cost Guides

AI Cost Optimization Guides & Best Practices

Practical techniques with explicit assumptions for measuring and reducing LLM API spend.

Prompt caching essentials

Reuse eligible stable prompts and reference context at the provider's cached-input rate, then measure the actual hit rate and cost change.

  • Place stable content (system prompt, tools, references) at the start.
  • Keep variable user input at the end of the prompt.
  • Enable caching on the static prefix and measure the hit rate.
Read the full guide

Model routing by difficulty

Send simple tasks to a small, cheap model and reserve flagship models for genuine reasoning to slash blended cost.

  • Classify each request by complexity before inference.
  • Route classification and extraction to a small model.
  • Escalate only hard tasks to the flagship model.

Output token control

Output costs several times more than input. Cap and shape responses to keep the most expensive tokens in check.

  • Set explicit max_tokens on every call.
  • Request concise or structured output where a machine consumes it.
  • Prefer long context with short answers over verbose generation.

RAG context efficiency

Retrieval quietly inflates input tokens. Tune chunking and reranking to send fewer, sharper passages.

  • Test several chunk sizes against a fixed evaluation set.
  • Rerank candidates and keep only passages that meet the measured relevance threshold.
  • Strip boilerplate and deduplicate before sending.
Read the full guide

AI agent budgeting

Agents re-send growing context on every loop. Cap steps and compress history to keep runs affordable.

  • Set a hard step limit per agent run.
  • Summarize conversation history instead of carrying full transcripts.
  • Use a cheap model for orchestration and tool selection.
Read the full guide

Infrastructure cost hygiene

Beyond tokens, retries, embeddings, and idle capacity add up. Make cost observable across the stack.

  • Reduce retries with robust error handling and validation.
  • Version embeddings and re-embed incrementally.
  • Right-size or autoscale self-hosted GPU capacity.