How accurate is this counter?

These are heuristic estimates based on each family's average characters-per-token ratio, not the exact tokenizer. For English prose they are typically within 5-10% of the real count; for code or non-Latin scripts the gap can be larger.

Why do different models show different counts?

Each model family uses a different tokenizer (BPE vocabulary). GPT models use tiktoken, Claude uses its own vocabulary, and Llama/Mistral use SentencePiece, so the same text splits into different numbers of tokens.

Is my text sent anywhere?

No. Everything runs in your browser. Nothing you paste is uploaded, stored, or logged.

How do tokens relate to words?

For typical English, one token is about three-quarters of a word, so 100 tokens is roughly 75 words. Punctuation, whitespace, and rare words split into more tokens.

Token Counter (Multi-model)

Multi-model token counter

Paste any text and instantly see how many tokens it is likely to use across the major model families — GPT (tiktoken), Claude, Llama, and Mistral — plus raw character and word counts. Useful for checking whether a prompt fits a context window or for estimating API cost before you send it.

How the estimate works

Exact token counts come from each model’s tokenizer, which we can’t bundle in a lightweight browser tool. Instead this counter uses each family’s measured average characters-per-token ratio and adjusts for whitespace and word boundaries. GPT models average roughly 4 characters per token for English; Claude and the SentencePiece-based Llama and Mistral families differ slightly, which is why the columns don’t match exactly. For English prose the estimates are usually within 5–10% of the real count.

Tips

Code, JSON, and non-Latin scripts tokenize less efficiently — expect more tokens than the English-tuned estimate.
For an exact count before a critical, high-volume call, run the provider’s own tokenizer (tiktoken for OpenAI, Anthropic’s count-tokens endpoint for Claude).
Remember both your prompt and the model’s reply count toward the context window — leave headroom for the output.