Exact token counts for DeepSeek, Qwen and GPT from each vendor's own tokenizer, not a character estimate. Paste text to compare five vocabularies side by side and see where every token splits. Free, runs in your browser.
| Model | Tokens | Tokens per character | Vocabulary |
|---|---|---|---|
| DeepSeek V4 / V3 / R1 DeepSeek · 128K | — | DeepSeek · 128K | |
| Qwen3.5 / 3.6 / 3.8 Qwen · 248K | — | Qwen · 248K | |
| Qwen3 / Qwen2.5 / QwQ Qwen · 151K | — | Qwen · 151K | |
| GPT-5 / GPT-4.1 / GPT-4o / o3 o200k_base | — | o200k_base | |
| GPT-4 / GPT-3.5 Turbo cl100k_base | — | cl100k_base |
These are the tokens of the text itself. An API call also counts the chat template around each message, so its reported prompt_tokens is usually a few higher.
Type or paste text to see every token.
Enter the input price from your provider's pricing page. Prices change, so the tool does not hard-code them.
Built and verified by the Go Tools Engineering Team.
About 4 characters of English per token It depends on the vocabulary and the language. Measured on this site's own text (about 82,000 Han characters of Chinese): English runs about 4.3–4.4 characters per token, or about 1.4 tokens per word; in Chinese prose a Han character costs about 0.8 tokens in DeepSeek and Qwen3.5, 0.9 in Qwen3, 1.0 in GPT-4o and 1.3 in GPT-4.
About 6,300 in DeepSeek, 10,000 in GPT-4 Counting punctuation as characters, measured on this site's Chinese prose: about 6,200–6,300 in DeepSeek, 6,400–6,500 in Qwen3.5, 7,000–7,200 in Qwen3, 7,700 in GPT-4o and 10,300–10,500 in GPT-4. Code and number-heavy text take more.
Same passage: 59–60 tokens in English, 71–156 in Chinese This page's “Chinese” and “English” examples are the same passage. The English version is 59–60 tokens in all five vocabularies; the Chinese one is 71 in Qwen3.5, 75 in DeepSeek, 84 in Qwen3, 111 in GPT-4o and 156 in GPT-4 — 18% to 160% more. English varies little between vocabularies (about 1.4 tokens per word on this site's text); for Chinese it depends on how many Chinese entries the vocabulary has, and one with few (GPT-4's cl100k_base) splits many characters into 2–3 byte-level tokens.
The smallest unit a model splits and bills text in A token is the smallest unit a language model splits text into, and the unit for API billing and context windows. It can be a word, part of a word, a character, a phrase or a few bytes of one character — each model's vocabulary decides.
No public tokenizer — only official counting APIs Not exactly. Claude has had no public tokenizer since Claude 3; Anthropic offers a count_tokens API instead (its documentation adds that Claude 4.7 and later use a newer tokenizer that produces about 30% more tokens for the same text). Gemini and Doubao are likewise counted through their official APIs, and Doubao also has an official token calculator in the Volcano Engine Ark console (sign-in required). This tool lists only models whose official vocabulary is public and matches id by id.
A token is the unit a language model reads and writes, and the unit APIs bill and context windows are measured in. A model never sees letters or Chinese characters directly: its tokenizer first cuts the text into pieces found in its vocabulary and turns each piece into an integer id. Common words and phrases have their own entries and take one token; anything missing falls back to UTF-8 bytes, so a single Chinese character can cost two or three tokens. The count therefore depends on the vocabulary. DeepSeek and Qwen store 人工智能 (artificial intelligence) as a single token and GPT-4o splits it in two, while GPT-4's cl100k_base has few Chinese entries and needs 12 tokens for the sentence below — more than twice DeepSeek's 5. Below is this tool's actual split.
文本：人工智能正在改变世界。（11 个字符） DeepSeek V4 5 [人工智能][正在][改变][世界][。] Qwen3.5 / 3.6 5 [人工智能][正在][改变][世界][。] Qwen3 / 2.5 5 [人工智能][正在][改变][世界][。] GPT-5 / GPT-4o 6 [人工][智能][正在][改变][世界][。] GPT-4 12 [人][工][智×2][能][正在][改][变][世×2][界][。] ×2 表示这个字被拆成了 2 个字节级 token
Qwen3.5, 3.6 and 3.8 switched to a new 248K vocabulary with different splitting rules; Qwen3, Qwen2.5 and QwQ keep the 151K one. Both come from the tokenizer.json in the official open weights and are counted separately — never approximated with a GPT tokenizer.
DeepSeek uses its official vocabulary (identical from V3 to V4.1); GPT uses OpenAI's o200k_base and cl100k_base. We compared 26,714 test texts (3.88 million characters across 15 languages, code, whitespace and characters at Unicode version boundaries) id by id against Hugging Face tokenizers and OpenAI tiktoken: all five vocabularies match.
One paste gives five counts, so you see at once which model handles your text with the fewest tokens. The tokens-per-character column compares texts of different lengths fairly.
The split view colors each token and can show its id. When several byte-level tokens only make whole characters together, they are grouped in one dashed block instead of showing garbled bytes; invisible characters such as zero-width spaces are written out by code point, e.g. ⟨U+200B⟩.
Opened files keep their original CRLF line breaks and BOM. DeepSeek splits a Windows line break \r\n into two tokens, so the same code with CRLF line endings costs about 15% more. Text pasted into the box is counted with the LF line breaks the browser keeps.
Tokenizing happens in a background thread (Web Worker); your text is not sent to our servers. Once the vocabularies are loaded, 200,000 Chinese characters are counted in under a second, and the text box stays responsive.
len(tiktoken.get_encoding("o200k_base").encode_ordinary(text)) OpenAI's official library, OpenAI vocabularies only. The most direct way to count GPT-5 and GPT-4o input tokens; it cannot count DeepSeek or Qwen.
AutoTokenizer.from_pretrained("Qwen/Qwen3.6-35B-A3B").encode(text, add_special_tokens=False) Loads the official tokenizers of open models such as DeepSeek and Qwen. The first run downloads the vocabulary files from Hugging Face.
The official offline tokenizer, used with transformers. Its vocabulary is the same as this tool's DeepSeek row.
messages/count_tokens, Gemini countTokens, OpenAI responses/input_tokens The only way for models without a public tokenizer, and the result includes chat-format overhead. It needs a request and an API key, and the text goes to the vendor.
No code and no dependencies — handy for checking one prompt or comparing models. For batch jobs, use the libraries above.
Click the “Chinese” example on this page: a 117-character everyday passage (106 Han characters)
DeepSeek V4 75 Qwen3.5 / 3.6 71 Qwen3 / 2.5 84 GPT-5 / GPT-4o 111 GPT-4 156
Per Han character: Qwen3.5 about 0.67, DeepSeek 0.71, Qwen3 0.79, GPT-4o 1.05 and GPT-4 1.47 tokens. Click “English” for the same passage in English: all five land between 59 and 60.
订单金额 12345.67 元，共 2026 件
DeepSeek V4 16 …[123][45][.][67]…[202][6]… Qwen3.5 22 …[1][2][3][4][5][.][6][7]…[2][0][2][6]… Qwen3 22 (digits split the same way) GPT-4o 13 …[123][45][.][67][ 元][，共]… GPT-4 16 …[123][45][.][67]…
Both Qwen generations split every digit on its own, while DeepSeek and GPT group digits in threes. Number-heavy text — tables, invoices, logs — costs noticeably more in Qwen.
你好世界 (three U+200B zero-width spaces between the characters)
Without them: DeepSeek 2 · Qwen3.5 2 · Qwen3 2 · GPT-4o 2 · GPT-4 5 With them: DeepSeek 7 · Qwen3.5 7 · Qwen3 7 · GPT-4o 7 · GPT-4 8
The zero-width spaces break 你好 and 世界 apart, so each character becomes its own token and each zero-width space takes one more. Text copied from the web often contains them; the split view shows them as ⟨U+200B⟩.
👍🎉
DeepSeek V4 3 Qwen3.5 6 Qwen3 2 GPT-4o 3 GPT-4 6
Qwen3's older vocabulary stores both emoji whole; the new 248K vocabulary drops them back to UTF-8 bytes, so a 4-byte emoji takes 3 tokens. A newer vocabulary is not cheaper everywhere — count your real text.
Paste a prompt, a document or some code, or use Open file to load a .txt, .md, .json, .csv or other text file. No text at hand? Click one of the examples. The first count downloads five vocabularies from this site (0.4–2.2 MB each, gzipped); they are not downloaded again while the page stays open.
The table shows the token count for DeepSeek, Qwen3.5/3.6/3.8, Qwen3/Qwen2.5, GPT-5/GPT-4o and GPT-4/GPT-3.5 at once, plus the average tokens per character. Above it you get characters, Han characters and UTF-8 bytes, to compare with character-based estimates.
Each colored block under “How the text is split” is one token. A dashed outline means several byte-level tokens only make whole characters together. Turn on “Show token IDs” to compare with the ids an API or tiktoken returns. Click any table row to switch models.
Enter the selected model's input price per million tokens to get the cost of this text as input. The tool ships no price list — vendors change prices often and split them by peak hours and cache hits, so use the number on their pricing page today.
A constant only holds for one kind of text. The same 24 characters 订单金额 12345.67 元，共 2026 件 are 13 tokens in GPT-4o and 22 in Qwen — the more digits, the bigger the gap.
tokens ≈ characters × 0.6
Split with the target model's vocabulary and count
tiktoken only ships OpenAI vocabularies. This page's “Chinese” example counts 156 tokens with cl100k_base, where DeepSeek uses 75 — an estimate twice too high.
tiktoken.get_encoding("cl100k_base") # used for DeepSeek AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4.1-Flash") The bill also includes the chat template, the system prompt, the output and a reasoning model's thinking; some services also price cache hits differently. Estimating from the input text alone comes out low.
cost = prompt tokens × input price
cost = prompt_tokens × input price + completion_tokens × output price
Zero-width spaces (U+200B) break words apart: 你好世界 is 2 tokens in DeepSeek and 7 once three zero-width spaces sit between the characters.
你\u200b好\u200b世\u200b界 → 7 tokens
你好世界 → 2 tokens
add_special_tokens=False to match this tool's plain-text count; DeepSeek's API documentation also offers an offline tokenizer package on its token usage page. Text Processing
Convert text between UPPERCASE, lowercase, Title Case, Sentence case, camelCase, PascalCase, snake_case, kebab-case, CONSTANT_CASE and 6 more formats instantly. Free, browser-only, no signup.
Text Processing
Generate Lorem Ipsum placeholder text instantly — by paragraph, sentence, word, byte, or list. Copy or download as plain text, HTML, Markdown, or JSON. 100% free, private, in-browser. No sign-up.
Text Processing
Test regex patterns instantly against any text. Live match highlighting, capture groups, replace preview, split, and pattern explainer. JavaScript-flavor regex, 100% private, no signup.
Text Processing
Remove duplicate lines from text online — free, instant, 100% in your browser. Ignore case or whitespace, keep first or last occurrence, and see how many duplicates were removed. Also sort and clean up. No signup.
Text Processing
Slugify any title into a clean, SEO-friendly URL slug instantly. Transliterate accents and Cyrillic, or keep Unicode letters. 100% private, in your browser.
Text Processing
Sort text lines alphabetically, numerically (natural order), or by length — ascending or descending. Free online line sorter, 100% in your browser. Also dedupe, reverse, and clean up lists. No signup, nothing uploaded.