Token Intelligence
Cross-model tokenization analysis for high-precision prompt engineering. Built with OpenAI’s official tokenizer, with labelled estimates for Anthropic and Google; everything is computed in your browser.
Input prompt
Characters
0
Words
0
OpenAI
GPT-4o
tokens
Anthropic
Claude
tokens
Gemini
tokens
Architect’s note
Different providers split words differently. Anthropic usually runs slightly lower than OpenAI for the same English text; Gemini has no public browser tokenizer, so its figure is a character estimate. For billing-critical numbers, call the provider’s own token-counting endpoint.
Frequently asked questions
Is my text uploaded when I count tokens?
No. Counting runs in your browser. The tokenizer code and vocabulary are downloaded to your device, and the text you type or paste is never sent to a server by this tool.
How accurate are the counts for each provider?
OpenAI counts are exact: they use js-tiktoken, a JavaScript port of OpenAI’s tiktoken, with the encoding that matches the model. Claude and Gemini counts are labelled estimates, shown with a “~” prefix, because neither provider publishes a tokenizer that runs reliably in a browser. Claude is estimated at roughly 2.7 characters per token and Gemini at roughly 4. For billing-critical numbers, use the provider’s own token-counting endpoint.
Why does the same text give different counts for different providers?
Each provider trains its own tokenizer with its own vocabulary, so the same sentence is split into different pieces. Differences are largest for code, non-English text and unusual whitespace, which is why counts from one provider should not be used to predict another’s.
Does the count include the extra tokens a chat request adds?
No. The tool counts the raw text you enter. A real chat request also carries per-message formatting and role markers, which add a small overhead on top. Treat the result as the size of your content, not the exact billed size of a full API request.