LLM Token Counter

An exact LLM token counter for GPT, with Claude and Gemini estimates

This LLM token counter splits your text the way OpenAI’s models do and gives the exact number of tokens for the two current OpenAI encodings: o200k_base (GPT-5.x, GPT-4.1, GPT-4o and the o-series) and cl100k_base (GPT-4, GPT-3.5 Turbo and the text-embedding-3 models). It also shows characters, words, lines and UTF-8 bytes, a colour-coded token visualiser, and bars that tell you whether a prompt fits the limits you type.

Claude and Gemini do not publish a tokenizer that runs offline, so their rows are clearly labelled estimates from each vendor’s own rule of thumb. For open models you can add the real Qwen3, Mistral NeMo and DeepSeek-V3.1 tokenizers. Paste a JSON array of chat messages and the page adds the per-message formatting tokens too.

Everything runs in a Web Worker in your browser, so even a long file stays responsive and nothing you paste is sent anywhere.

How to count tokens

1. Paste text into Text, prompt or chat messages, or use Open a file for .txt, .md, .json, .csv, code or the text of a PDF.
2. Read the exact o200k_base and cl100k_base counts, and the Claude and Gemini estimates, in the table.
3. Under Context fit, type your model’s limit and how many tokens to keep free for the answer.
4. Use the Token visualiser to see where the text is split; hover a token for its id, or tick Show token ids.
5. For an open model, open Add open-model tokenizers and load Qwen3, Mistral NeMo or DeepSeek-V3.1.

LLM Token Counter features

What a token is, with real numbers

A token is a chunk of text from the model’s vocabulary — often a whole common word with its leading space, sometimes a fragment. OpenAI’s rule of thumb is that one token is about four characters or three quarters of a word of English, but the real split depends on the text. These counts come from this tool’s tokenizer and match the examples in OpenAI’s Cookbook:

The larger o200k_base vocabulary helps most outside English: the Hindi greeting needs 13 tokens in the older encoding and 5 in the newer one. To turn counts into money, use the OpenAI token calculator.

Which models use which tokenizer

OpenAI’s tiktoken library (model table checked 14 October 2026) maps GPT-4o, GPT-4.1, GPT-4.5, o1, o3, o4-mini and every “gpt-5…” model name to o200k_base, and GPT-4, GPT-3.5 Turbo and the text-embedding-3 models to cl100k_base. OpenAI has not published the encoding of the GPT-6 models; for those, o200k_base is the closest figure available offline, and OpenAI’s token-counting endpoint gives the exact number.

Anthropic says a token is about 3.5 English characters on models before Claude Opus 4.7, and that Opus 4.7 and later use a tokenizer that produces about 30 % more tokens for the same text. Google says a Gemini token is about 4 characters. Those ratios are averages for English: short, technical or non-English text can be far off, so use the vendors’ token-count APIs when an exact figure matters.

Chat messages cost more than their text

Every message carries hidden formatting — the role, an optional name and markers between messages. OpenAI’s Cookbook counts 3 tokens per message, 1 more for a name and 3 to prime the reply, and its example conversation comes to 129 tokens in cl100k_base and 124 in o200k_base, the same totals OpenAI’s API reported. Paste that conversation here and you get the same numbers. OpenAI calls the formula an estimate for newer models; tools, images and files add tokens that only the API can count.

Tips for staying under a limit

• Leave room for the answer: the limit covers input and output together, and reasoning models also spend hidden reasoning tokens.
• Trim repeated instructions and filler with the prompt optimizer — it shows the exact tokens saved.
• Structure long prompts with the prompt formatter so you can see which part is longest.
• Words still matter for people: the word counter shows reading time.