Six levers to cut your LLM API bill

Switching to a cheaper model is the obvious move, and often the wrong first one. Here are six levers, roughly in order of effort-to-payoff.

1. Right-size the model

Most production traffic does not need a frontier model. Route the easy 80% to a small model and reserve the expensive one for the hard cases. A blended cost of $0.60/1M instead of $6.00/1M is a 10× cut with no quality loss on the easy path.

2. Prompt caching

If your system prompt or retrieved context repeats across requests, prompt caching cuts the input price on the cached portion by 75-90%. For RAG and agent workloads where the prompt dwarfs the completion, this is the single biggest lever.

3. Batch the non-urgent work

Evals, back-fills, classification jobs, nightly summaries - anything that does not need an answer in seconds - belongs on the batch API, usually at half price.

4. Trim the prompt

Few-shot examples you no longer need, verbose instructions, whole documents where a snippet would do. Input tokens are cheaper than output but they are not free, and long prompts also slow you down.

5. Cap the output

Output tokens cost 3-5× input. Set max_tokens, ask for terse answers, use structured output instead of prose where you can.

6. Then compare providers

Once the workload is lean, the same model is often available cheaper via an aggregator or a different host. That is the last 10-30%, not the first.