Estimate your bill before you build

A repeatable method to budget an LLM feature and keep costs under control.

You can estimate an LLM bill with five numbers and a calculator. Do it before you write the feature.

Five steps

1. Measure a typical request. Count input tokens (instructions, documents, history) and output tokens for a realistic example. Providers offer token counters, and ours has rules of thumb.
2. Estimate volume. Requests per day times 30. Be pessimistic.
3. Account for caching. What share of the input is the same long prefix each time? That part can be billed at the cached rate.
4. Add a margin for retries. Failed or repeated calls cost money. Ten to twenty percent is a reasonable start.
5. Compare models. Enter the numbers in the calculator and look at the cheapest few. Test whether they are good enough on your own examples.

Keep it under control

• Set a maximum output length for every request.
• Trim conversation history and send only what the model needs.
• Use a smaller model for easy tasks and a stronger one only when needed.
• Set budget alerts and spending limits in your provider account.
• Handle rate limit errors (HTTP 429) with exponential backoff instead of tight retry loops.

Common surprises

• Chat history grows each turn, so cost per message rises over a conversation.
• Long system prompts are paid on every request unless cached.
• An agent that loops can use thousands of tokens before a human notices.

NoteCheapest is not best. Evaluate quality on your own data before you choose a model for cost alone.