Chinese LLM Token & Cost Calculator

Tokens, not words, are how LLMs bill — and Chinese counts differently

Every major large language model bills by the token, and a token is not the same as a word or a character. Roughly, a token is a short chunk of text — in English about four characters on average, but Chinese is far denser, where a single character can be close to one or two tokens on its own. That is why you cannot price an LLM call by counting words: the only honest unit is the token. This calculator gives you two ways to get there. Paste your prompt, document or conversation and it estimates the token count using a tokenizer that runs entirely in your browser; or, if you already have exact figures from your API usage logs, switch to manual mode and type the input and output token counts straight in. Either way, it then multiplies those tokens by each model's published input and output rates and ranks the results from cheapest to most expensive, so the trade-offs are visible at a glance.

The reason input and output are entered separately is that almost every provider prices them separately — and output is usually the more expensive of the two, often by several times. The text you paste represents only the input you send; the length of the model's reply, the output tokens, is something you have to estimate yourself before the request is made. Reasoning models complicate this further: a model like DeepSeek R1 generates a large volume of intermediate "thinking" tokens that are typically billed as output, so even a short final answer can carry a surprisingly high per-call cost. When you are budgeting for one of those, set the output token field higher than the visible answer length would suggest, or your estimate will run low.

Estimates, not invoices: read the numbers as planning figures

It is important to be clear about what this tool can and cannot tell you. The token counts are estimates, not exact figures. They are produced with OpenAI's o200k_base tokenizer because it is the one that can run locally and openly in a browser — but DeepSeek, Qwen, Kimi, Zhipu and the rest each use their own tokenizer, and the same text will segment into a slightly different number of tokens on each. Treat the count as a sound planning figure, accurate enough to compare models and size a budget, but not as the exact number that will appear on your invoice. The same caution applies to the prices: the rates here are provisional, gathered from public aggregators, and the official price from each provider always prevails. Models whose rate has not yet been verified are listed by name with a "—／待核实" marker and left out of the arithmetic rather than guessed at.

Used with that understanding, the calculator is genuinely useful for the decisions that matter before you spend money. Run the same workload across the models and the cost gap is frequently several to more than ten times, which can change which model you choose, how long you let replies run, and whether a feature is viable at scale. Multiply a single request by your expected monthly volume and you have a defensible budget line before a single API key is issued. And because the whole thing runs locally — no upload, no model call, no stored input — you can iterate as much as you like: paste a real prompt, read the spread, tighten the output length, switch models, and re-check, all without anything you type ever leaving your device. When the decision is final, confirm the current rate on each provider's own pricing page and treat the figure here as the starting estimate it is meant to be.