Chinese Token Budgets Need More Than Character Counts

Chinese character counts are only a starting point

A Chinese document containing a thousand characters does not necessarily contain a thousand tokens. A token is a unit a model uses to process text, and its boundaries come from the model's tokenizer rather than ordinary writing rules. A token can represent a character, part of a character, or a sequence of characters. Spaces are not required for token boundaries.

This makes character counts useful for measuring document size, but unreliable as a universal billing estimate. The familiar English shortcut of roughly four characters per token should not be carried over to Chinese. Even within Chinese text, vocabulary, punctuation, and embedded English can change the relationship.

The selected model determines the useful counting method

Different models can tokenize the same passage differently. A tokenizer designed with substantial Chinese coverage may encode familiar sequences efficiently, while another may split them into more units. Rare characters, unusual symbols, and specialized terminology can increase counts. A paragraph mixing Chinese prose with product identifiers or numbers may behave differently from an equally long literary passage.

Use the exact model's counting tools when available. Consult its provider documentation for a supported tokenizer or token-counting endpoint. A counter intended for another model family is only a rough proxy. Tokenizer versions and request formats also matter, so record the model and counting method alongside any estimate you plan to reuse.

Representative samples produce better estimates than universal ratios

Start with several passages resembling the material you will actually send. Include ordinary paragraphs, headings, lists, and technical vocabulary in approximately their expected proportions. Count their characters consistently, then measure their tokens with the relevant tool. Divide tokens by characters to obtain a working ratio for that material.

For example, suppose a sample containing 600 characters measures 780 tokens. Its observed ratio is 1.3 tokens per character. A similar document containing 2,000 characters would then suggest about 2,600 text tokens. These numbers illustrate the method, not a standard Chinese conversion rate. Repeat the measurement across several samples, especially when documents vary substantially.

For planning, keep both a typical ratio and a higher observed ratio. Use the typical value for forecasting ordinary traffic and the higher value when checking capacity. Recalibrate after changing models or moving from one content type to another.

The full request contains more than document text

Your document is only part of the input. System instructions, conversation history, tool definitions, and other request elements can consume tokens too. Message formatting may introduce additional overhead. Counting a pasted paragraph alone therefore does not establish the token count of the complete request.

Leave space for the response as well. Input and generated output generally share the model's context capacity, although exact limits and accounting rules depend on the provider. Some models also have reasoning-related usage categories. Check the selected model's documentation rather than assuming every consumed token appears in the visible answer.

Provider usage reports turn estimates into billing evidence

Estimation supports budgeting; reported usage supports reconciliation. BestRelay, which publishes this article, is a quality-first API relay that forwards each request to the official model named in it, without substitution. Its coverage includes Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, and Grok. Keeping the requested model explicit helps keep token comparisons meaningful.

BestRelay bills pay-as-you-go using provider-reported token counts, with cache reads and writes priced separately. Separate those categories when projecting costs: an unchanged token count does not imply an unchanged bill when caching treatment differs. Members can also earn referral commission when invited users top up; keep that income separate from usage-cost calculations.

A small measurement routine keeps Chinese budgets dependable

Measure representative inputs, include request overhead, reserve output capacity, and compare estimates with actual usage after deployment. BestRelay works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK, but client compatibility does not make tokenization identical across models. Create a BestRelay account to test a representative Chinese workload. Treat each estimate as a measured approximation, not a permanent property of Chinese writing.