How to estimate a monthly token bill

Send a typical request through the model you plan to use. Read the input and output token counts in the response. Do it for each kind of request your application makes: a short question, a long document, an agent step.

Estimate how many requests of each kind you will send in a month. Multiply the counts. You now have a month of input tokens and a month of output tokens, in millions.

Be honest about growth inside a conversation. The tenth turn of a chat resends the nine before it, so its input is far larger than the first.

Multiply the input millions by the input price and the output millions by the output price. On 12 Nov 2026, the lowest listed input price for Qwen3.8 27B here was $0.035 per million tokens. The lowest listed output price was $0.395 per million tokens.

A month that reads two hundred million tokens and writes twenty million costs two hundred times the first figure plus twenty times the second. Take both prices from the same host, since the two lowest figures may sit in different rows.

If a long opening repeats, price that part at the cached input rate. Then add a margin for retries, failed requests that were billed, and test traffic. These are real tokens and they are on the bill.

Run the application for a day and read the charge. Compare it with a thirtieth of your estimate. If they differ, the token counts are wrong, not the prices. Find the request that is larger than you thought.

Runcrate bills its 125 models from one prepaid balance, so a day of testing has a clear cost and cannot run past what you loaded. Runcrate serves Qwen3.8 27B at $0.158 per million input tokens and $1.97 per million output tokens.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.