Skip to content
Yunedia Tools

Contador de tokens de LLM

Conte tokens com o tokenizador BPE real para modelos da OpenAI (exato), além de estimativas claramente identificadas para Claude, Gemini e outros modelos.

Enter some text to see the count.

Como usar

  1. Paste the text you want to count.
  2. Pick a model. GPT-4o family and GPT-4/3.5 are counted exactly with the real tokenizer.
  3. Claude, Gemini, Llama and Mistral show an estimate, always labeled as one, with the heuristic stated.

Observações

Exact counts run entirely in your browser with gpt-tokenizer (o200k_base / cl100k_base); your text is never sent to a server. Other providers do not publish a browser tokenizer, so this page offers only a character-based estimate for them and never presents it as exact.

Perguntas frequentes

How far off can the estimates be?
For English prose, often within 10–20%, but code, long numbers and CJK text can be far off. Keep a margin when budgeting from an estimate.
Why does Japanese cost more tokens?
BPE tokenizers are trained mostly on English, so a Japanese character typically becomes about one token (sometimes more). You can verify this with the exact GPT counters on this page.