AI Cost Calculator

The same playbook covers OpenAI API cost and the equivalent bill from Anthropic or any other provider. Start with the three levers that need no contract change: route cheap requests to a smaller model, cache the repeated part of your prompts, and move asynchronous work to a batch endpoint. Then trim dead context. Most production prompts carry instructions and examples that no longer earn their tokens. Volume discounts and committed-use deals are worth asking about, but they come last, because otherwise you negotiate a discount on spend you could have removed.

Prompt caching stores the unchanging prefix of a request (system prompt, tool definitions, a long document) on the provider's side, so repeat calls are billed at a reduced rate for those tokens instead of the full input price. It pays off when many requests share a long prefix and that prefix is byte-identical, so put the stable part first and the variable part last. Discount size, minimum prefix length, and cache lifetime differ by provider; check the current terms before assuming a number.

A batch endpoint takes a file of requests and returns the results within a stated window, commonly up to 24 hours, in exchange for a lower price per token. It fits work with no user waiting on it: data enrichment, evaluation runs, backfills, classifying a queue, nightly summaries. It does not fit interactive chat or anything in a checkout path, and batch requests usually cannot be combined with every other discount, so read the terms.

When the task is narrow and the output is short or structured: classification, routing, extraction into a schema, tagging, short rewrites, first-pass drafts. Decide it by measurement: collect a few hundred real requests, score both models against them, and route on the result rather than on intuition. Keep the frontier model for reasoning, long context, and anything a customer reads word for word.