LLM cost calculator

Two numbers decide an LLM bill: how many tokens go in, and how many come out. Input tokens are the ones teams underestimate — a system prompt, retrieved context and conversation history are resent on every single call, so a chat feature can spend ten times more on input than on the answer it produces.

Fill in your own numbers below. The prices are editable because provider pricing moves and a calculator that quotes last year's rates is worse than none at all; the defaults are indicative figures as of September 2026, and you should overwrite them with what your invoice actually says.

What the crossover means

The comparison at the bottom is the one worth taking seriously. A metered API costs nothing when idle and scales linearly with use; a dedicated instance costs the same every month whatever happens. Below the crossover the API is cheaper, above it the instance is.

The catch is that the two are not equal in capability. A quantized 7-8B model on CPU is a capable classifier, extractor and summariser, and it is not a frontier model. The honest comparison is between the API calls you would replace — the high-volume, low-difficulty ones — and the instance, not between your whole bill and one server. Our CPU performance figures set out which jobs fit.

Where the money usually hides

Retries. SDKs, queues and workflow engines all retry. A failing call can cost three times what the logs suggest.

Agent loops. Plan, act, observe, repeat — without a step cap, one request can become forty calls.

Context that grew. System prompts accumulate. Nobody deletes anything from them.

A gateway with per-key spend logging answers all three in a week. We host LiteLLM from $5.99/mo, and the seven levers guide covers what to do once you can see the numbers.