LLM Token Counter
Count the tokens in a prompt with each model's own tokenizer, and compare the count across models.
153 characters · 26 words · Counted in your browser
Tokens
…
Compare across different models
…
How the token count is made
Each model is counted with the tokenizer its creator published, not with a characters-per-token estimate.
The model's own tokenizer
The file each creator published with the model, run in your browser. The count is exact for the text.
Checked per model
A model is listed only after its tokenizer matched on 190 texts in 20 languages, or came within a tenth of a percent.
Text only
A chat request adds a few tokens per message. For the background, read what tokenization is.
Supported models and tokenizers
41 models, grouped by the tokenizer they share.
- DeepSeek tokenizer
- DeepSeek V4 FlashDeepSeek V4 Flash 0731DeepSeek V4 ProDeepSeek V3.2DeepSeek V3.1DeepSeek V3DeepSeek R1 0528Step 5 Preview
- Qwen3 tokenizer
- Qwen 2.5 72B InstructQwen 3 14BQwen 3 30B A3BQwen 3 32BQwen3 Coder 480B A35B InstructQwen3 Next 80B A3B InstructQwen 3 VL 235B A22B InstructQwen 3 VL 30B A3B InstructBeam 1.0 501B A23B
- Qwen3.5 and later tokenizer
- Qwen 3.5 9BQwen 3.5 27BQwen 3.5 35B A3BQwen 3.5 122B A10BQwen 3.5 397B A17BQwen 3.6 27BQwen 3.6 35B A3BQwen 3.8 2.4T A95B
- OpenAI o200k tokenizer
- GPT OSS 120BGPT OSS 20B
Token count by language
One sentence, written in each language and counted on each tokenizer. An example, not an average: pick a sample above to see the text.
| Language | Characters | DeepSeek | Qwen3 | Qwen3.5 and later | OpenAI o200k | GLM | Mistral |
|---|---|---|---|---|---|---|---|
| English | 153 | 30 | 29 | 29 | 29 | 29 | 29 |
| Spanish | 168 | 39 | 39 | 37 | 35 | 38 | 36 |
| German | 180 | 43 | 43 | 36 | 33 | 38 | 38 |
| Russian | 170 | 46 | 51 | 41 | 43 | 42 | 47 |
| Chinese | 52 | 28 | 31 | 30 | 35 | 30 | 47 |
| Japanese | 81 | 50 | 48 | 35 | 57 | 48 | 58 |
| Korean | 80 | 54 | 60 | 45 | 48 | 67 | 46 |
| Hindi | 164 | 92 | 151 | 67 | 49 | 165 | 97 |
| Arabic | 121 | 50 | 53 | 40 | 43 | 56 | 40 |
| Python code | 157 | 38 | 35 | 38 | 35 | 35 | 35 |
Tokens per tokenizer. The fewest in each row is highlighted.
Explore more
Companion tools that draw on the same catalog and routing engine.
Frequently asked questions about token counting
How tokens are counted, why models disagree, and what the count means for cost and context.
Paste the text, pick the model, and read the count. The page runs the tokenizer that model's creator published, so the number is the model's own count of the text, not a rule of thumb. Counting happens in your browser as you type.
A tokenizer cuts text into tokens, the pieces a language model reads and writes. Each model ships with its own tokenizer and a fixed list of pieces, so a token can be a whole word, part of a word, a single character or a punctuation mark. Context windows, output limits and prices are all measured in tokens.
For English prose the common rule of thumb is about 1,300 tokens, or three quarters of a word per token. The real figure depends on the model and the text: code, numbers and languages such as Chinese or Japanese take more tokens for the same length. Paste the text and pick the model to get the exact count.
Each creator trains its own tokenizer on its own data, so the same text is cut into different pieces. Common English words are usually one token everywhere, and the differences grow with other languages, code and unusual formatting. The comparison table on this page shows the count for every group of models side by side.
It is close for English and wrong for much else. In a measurement of 7,990 chat texts in 14 languages (November 2026), dividing characters by four undercounted Chinese by a little over two times at the median and Japanese by about 1.8 times, and it undercounted 28.5% of all texts by more than a quarter. Counting with the model's real tokenizer removes most of that error.
41 models across 6 tokenizers: DeepSeek, Qwen3, Qwen3.5 and later, OpenAI o200k, GLM, Mistral. A model is listed only after its creator's own tokenizer gave the same counts as the file used here on 190 test texts. Models whose tokenizer is not published, or whose license does not allow the file to be redistributed, are not listed.
No. The tokenizer file is downloaded to your browser and the counting runs there, in a background worker. The text you paste is not sent to Inferbase or to anyone else.
It matches the model's count of the text you paste. A provider's bill for a chat request is a little higher, because the chat template adds a few tokens per message for role markers, and tool definitions and images are counted too. Treat the number here as the text's own size.
A model's context window is measured in tokens and covers the prompt and the answer together. The page shows what share of the selected model's window the text fills, so you can see whether a document fits and how much room is left for the answer.
Multiply the token count by the price per token. Enter a price per million input tokens and the page does the multiplication for the selected model. Output tokens are priced separately and usually cost more than input tokens.
Some models in a family use a tokenizer that differs from the shipped file by a handful of entries. Where the totals agree within a tenth of a percent, the model is listed with that figure beside it. Where they differ by more, the model is not listed.
Put intelligence in the middle.
Start on the free plan with 5,000 requests a month, no card required.