Tokens and token pricing

What is a token?

A token is a unit produced by a model's tokenizer when it converts text into a form the model can process. It may represent part of a word, a whole word, punctuation or another sequence. Tokens are not a fixed number of characters, and the same sentence can be split differently by different tokenizers.

Input tokens describe material supplied to the model, while output tokens describe material it generates. The input may include more than the user's latest message: instructions, conversation history, retrieved documents and tool results can also contribute. A realistic estimate needs to account for that whole request.

How does per-token pricing work?

A per-token service applies a rate to the measured input and output usage. When rates are quoted per million tokens, divide each token total by one million, multiply it by the corresponding rate, and add the two amounts. Use the separate rates rather than applying one blended price to the entire exchange.

IndicStack publishes the current Qwen3.8 27B input and output rates on its pricing page, before GST. The same page includes a worked example and the Unlimited Pro and Unlimited Team packages. Consult those figures when estimating a workload rather than copying a rate into a document that may remain unchanged after prices are updated.

A short prompt can still produce a lengthy answer. Conversely, a large document request may have substantial input usage even if the answer is brief. Measure representative interactions so both sides of the calculation reflect your application.

How do I estimate a monthly workload?

Group requests by their purpose, such as document summarisation, extraction or multi-turn assistance. Measure typical input and output usage for each group and multiply by the expected request volume. Keep unusual large requests visible in the estimate, rather than assuming all calls look like the average.

Conversation design changes usage. Repeatedly sending an entire history, adding irrelevant retrieved passages or returning unnecessarily long answers can increase the token total. Review those decisions against answer quality before reducing context or shortening responses.

Separate token costs from other application costs, including storage, retrieval and tools. For agency work, track spend by client key so forecasts can be compared with actual usage. That comparison helps explain changes in a bill and provides a better basis for the next estimate.

Related terms