Free GPT Token Counter & OpenAI Tokenizer

A token is the smallest unit of text that a language model processes. OpenAI's tokenizers split text using Byte Pair Encoding (BPE), where a token can be a whole word, part of a word, a single character, or whitespace. As a rough guide for English text: 1 token ≈ 4 characters 1 token ≈ 0.75 words 100 tokens ≈ 75 words 1,000 tokens ≈ 750 words (about 1.5 pages)

This tool uses the official gpt-tokenizer library with the exact BPE vocabularies OpenAI ships: o200k_base — GPT-5, GPT-5 mini, GPT-5 nano, GPT-4.1, GPT-4.1 mini/nano, GPT-4o, GPT-4o mini, o1, o3, o4-mini cl100k_base — GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, text-embedding-3-large, text-embedding-3-small Token counts produced here match what the OpenAI API bills you for.

Yes — for plain text input, the count is identical to what the OpenAI API and Playground report. The tool uses the same open-source BPE vocabulary that powers tiktoken. Note: chat completions also add a small constant overhead (roughly 3–4 tokens per message for role and separator tokens). For precise chat-message accounting, add ~4 tokens per message on top of the counted content.

Cost estimates use OpenAI's published per-million-token prices (updated November 2025). The input cost multiplies your measured tokens by the model's input price; the output cost is based on the number of output tokens you expect. Always double-check OpenAI's pricing page before committing to a large job, since rates occasionally change.

No. All tokenization happens entirely in your browser using WebAssembly-free JavaScript. Your prompts, documents, and source code never touch our servers, are never logged, and never leave your device.

Each model family uses a different BPE vocabulary. Newer o200k_base models (GPT-5, GPT-4o, GPT-4.1, o-series) were trained on a larger, more efficient vocabulary, so they typically produce 10–15% fewer tokens than older cl100k_base models (GPT-4, GPT-3.5), especially for code, emoji, and non-English text.

BPE vocabularies are trained primarily on English. Non-English characters — especially CJK (Chinese, Japanese, Korean) and right-to-left scripts — often fall back to multiple sub-byte tokens, consuming 2–4× more tokens per character than English equivalents. The newer o200k_base tokenizer significantly improves this, which is why it's recommended for multilingual apps.

Common ways to shrink a prompt without losing meaning: Remove redundant instructions and examples. Use shorter variable and placeholder names. Replace Markdown bullet lists with comma-separated lines when possible. Summarize long context with a cheaper model before feeding it to an expensive one. Use newer models (GPT-4o, GPT-4.1, GPT-5) — their tokenizer is more efficient.

Yes. Pick text-embedding-3-large or text-embedding-3-small from the model dropdown. You'll see the input token count and the exact price per request — embeddings have no output cost.

Yes — the GPT Token Counter is completely free, with no rate limits, no sign-up, and no API key required. Tokenization runs locally in your browser.