LLM API Cost Calculator

The spread between the cheapest and most expensive model for the same workload is routinely 50x or more. Before spending a week shaving tokens off a prompt, check whether a smaller model handles the task at all, because that single decision moves the bill further than any amount of prompt engineering.

The reverse also holds. Routing everything to the cheapest model to save money often costs more overall, because a weaker model needs more retries, longer prompts, and more human correction. The useful comparison is cost per successfully completed task, not cost per million tokens.

A common structure is to route by difficulty: a small model handles the bulk of routine requests, and a frontier model handles the remainder that the small one cannot. Model the two tiers separately here and add them together.

Output is far more expensive than input, typically by 5x, and reasoning tokens are billed as output on models that produce them. A request with 100 visible output tokens can bill for thousands if the model reasoned at length first. If your workload uses extended or adaptive thinking, raise the output figure well above what you see in the response.

System prompts are charged on every request. A 4,000-token system prompt across a million requests is four billion input tokens, which is a meaningful line item on its own. This is exactly the content prompt caching exists to make cheap, which is why the cache slider is on this page rather than hidden behind an advanced toggle.

Retries and failures still bill for whatever was generated before they failed. A pipeline with a 5 percent retry rate is 5 percent more expensive than a naive calculation suggests, and rate-limit backoff loops can be considerably worse if they are retrying long prompts.

The pricing table ships with the page and all arithmetic happens locally. Nothing about your workload, volumes, or token counts is uploaded, which matters because request volumes and prompt sizes are commercially sensitive information that has no business being typed into someone else's server.