Token Counter

It depends on the model, and every count is labeled. GPT-5 series counts are exact: they use o200k_base, the tokenizer OpenAI's tiktoken library assigns to GPT-5 models, and match tiktoken token for token. Every other count is an estimate built from the provider's own published ratio where one exists. The table under the counter shows which is which.

Here you get an estimate, because Anthropic doesn't publish Claude's tokenizer. Exact counts come from Anthropic's token counting API, which is free to call but needs an API key. The estimate uses Anthropic's figure of about 3.5 English characters per token. Claude Opus 4.7 and later, including Opus 5.5, Fable 5.1 and Sonnet 5, use a newer tokenizer that Anthropic says gives about 30% more tokens for the same text, so don't reuse counts from Haiku 4.5 or Sonnet 4.6.

OpenAI hasn't said. As of September 2026, tiktoken maps GPT-5 models to o200k_base but has no entry for GPT-6 Astra, Sol or Luna. This tool counts GPT-6 with o200k_base and labels the result an estimate. If GPT-6 turns out to use a new tokenizer, real counts could differ.

From each provider's published ratio. Google says a Gemini token is about 4 characters. DeepSeek says an English character is about 0.3 tokens and a Chinese character about 0.6. Qwen's docs give 3 to 4 English characters, or 1.5 to 1.8 Chinese characters, per token, and we use the midpoints. Grok, Kimi, GLM, MiniMax, MiMo and Muse Spark have no published ratio, so their estimate averages characters divided by 4 and words times 1.3.

A chunk of text the model reads as one unit: a word, part of a word, a space plus a word, or a punctuation mark. Common English words are usually one token each. Code, numbers, rare words and non-English text break into more pieces. The token view shows exactly where o200k_base splits your text, with each token in its own color.

Chat APIs wrap each message in a few formatting tokens, and system prompts, tool definitions and images count as input too. This tool counts the text you paste, so expect a handful of extra tokens per message on a real request.

Tokens times the model's list price per million input tokens. The cost shown covers input only. Output and thinking tokens are billed separately, usually at a higher rate. For a whole job, the cost per 1,000 prompts is the more useful number.

The most tokens a model can handle in one request, counting your input and its reply. The bar shows how much of the selected model's window your text fills, so you know before sending whether a long document fits.

Yes. Counting runs in your browser. The GPT tokenizer loads from this site the first time you type, and your text is never sent anywhere. There's no signup.