Five LLM cost levers to check before you switch providers

When an AI bill grows, the first idea is often to switch providers. A migration costs weeks and new quality risks, and most of the waste stays behind: it comes from how models are used, not whose models they are. These five levers reduce LLM costs with the provider you already have. Prices below are from the providers' own pages, checked November 7, 2026.

1. Route each task to the right model

A ticket classifier and a contract review do not need the same model. Per input token, the most expensive model on Anthropic's price list costs 100 times the cheapest, and OpenAI's list has the same gap (Anthropic, OpenAI). Test cheaper models against real examples per task and keep the cheapest that passes: right-modeling.

Measure: cost per request on each route.

2. Cache the parts of the prompt that repeat

System prompts, tool definitions and shared documents are sent on every call. A cached prompt read costs 10% of the normal input price on most Claude models (Anthropic). Put the stable parts first and the variable parts last.

Measure: the share of input tokens billed as cached.

3. Move work that can wait to batch

Overnight reports, bulk tagging and evaluation runs can wait hours. Batch requests cost half the standard price on Anthropic, OpenAI and the Gemini API (OpenAI, Google), and Amazon Bedrock offers 50% off on-demand for selected models (AWS).

Measure: the share of tokens processed in batch.

4. Limit and shape the output

Output tokens cost five times input tokens on the flagship models of Anthropic, OpenAI and Google (pricing pages above). If the interface shows two sentences, ask for two sentences, and cap the output per route.

Measure: average output tokens per request.

5. Send less context

Send the passages that matter, summarize long histories, and check whether quality changes when you halve the retrieved chunks.

Measure: average input tokens per request.

Make each lever a change you can review

Each lever is one small change with an estimate before and a measurement after: the same way we handle cloud savings in Savings as Code. Then write the chosen model and limits down where they are checked, so the next feature does not undo them: Spend as Code, coming soon.

If the bill is still too high after that, compare providers. You will be comparing the real workload, not the waste.