--- title: "Token Counter for GPT-5, Claude, Gemini & more" description: "Free token counter for GPT-5, Claude, Gemini, DeepSeek, and more. Real tokenizers run in your browser, nothing uploaded, with cost estimates as you type." updated: 2026-08-13 canonical: https://holaclaw.ai/tools/token-counter --- # Token Counter and Cost Estimation Paste any text and see how it tokenizes across model families, with live token counts, cost estimates, and context-window usage. The tool runs each family's real tokenizer in your browser: nothing is uploaded, and there is no sign-up. Supported families: GPT (o200k_base, cl100k_base, o200k_harmony), Claude, Gemini, Gemma, Muse Glimmer, DeepSeek, Qwen, Mistral, Grok, and GLM. Use it at [holaclaw.ai/tools/token-counter](https://holaclaw.ai/tools/token-counter). ## What is a token? **A token is the unit of text a language model reads: a short chunk of characters, often a whole word, sometimes only a piece of one.** Models like GPT-5, Claude, and Gemini never see letters or words directly. A tokenizer first splits your text into pieces from a fixed vocabulary and turns each piece into a number. Common English words usually map to a single token, while rarer words, names, code, and text in other languages split into several. Providers bill their APIs per token and measure context windows in tokens, which is why the count matters. ## Why does the same text give different token counts per model? **Each model family uses its own tokenizer with its own vocabulary, so the same text splits into different pieces.** OpenAI's o200k_base vocabulary has about 200,000 entries; other families made different choices about which character sequences deserve an entry of their own. This tool runs the real tokenizer for each family directly in your browser, so the differences you see are actual splits, not estimates. On plain English prose the spread stays small: our fixed 500-word sample ranges from 571 tokens with Grok's tokenizer to 581 with Mistral's. It widens on code, non-English text, and unusual formatting, where vocabularies differ more. Two exceptions: Anthropic and Google do not publish the tokenizers of current Claude and Gemini models, so those counts are close approximations and the tool marks them as estimates. ## How does the cost estimate work? **The estimate multiplies your token count by the model's public price per million input tokens.** For example, 10,000 tokens on a model priced at $1.25 per million input tokens comes to $0.0125. The estimate covers input tokens only; model replies are billed as output tokens, usually at a higher rate. Open-weight models show as free because you can run them on your own hardware. Prices are refreshed periodically from providers and OpenRouter, so treat the figure as guidance rather than a quote. ## How many tokens is 500 words? **In our measured 500-word English sample, every family lands between 571 and 581 tokens, close to 1.15 tokens per word.** The table below works as a quick word to token calculator for planning. For exact numbers, paste your own text into the token count estimator: it runs each family's real tokenizer, so you don't have to rely on averages. | Model family | Tokens for 500 words | Tokens per word | Characters per token | | --- | --- | --- | --- | | GPT-5 / GPT-4o (o200k_base) | 572 | 1.14 | 4.92 | | GPT-4 Turbo / GPT-3.5 (cl100k_base) | 574 | 1.15 | 4.90 | | Claude (claude-v4) * | 574 | 1.15 | 4.90 | | Gemini (gemma-3 proxy) * | 578 | 1.16 | 4.87 | | Gemma (gemma-4) | 578 | 1.16 | 4.87 | | Muse Glimmer (muse-glimmer-30b) | 578 | 1.16 | 4.87 | | DeepSeek (deepseek-v4) | 572 | 1.14 | 4.92 | | Qwen (qwen-3.6) | 579 | 1.16 | 4.86 | | Mistral (mistral-3-4) | 581 | 1.16 | 4.84 | | Grok (grok-2) | 571 | 1.14 | 4.92 | | GLM (glm-4.6) | 574 | 1.15 | 4.90 | Measured with each family's tokenizer running in this tool, on a fixed 500-word English prose sample (2,812 characters). Ratios shift with code, other languages, and formatting. Entries marked with an asterisk use a close approximation because the maker does not publish the tokenizer. ## Token counters by model ### Claude token counter Select any Claude model in the tool to count tokens for Claude Opus, Sonnet, or Haiku. Anthropic does not publish the tokenizer for current Claude models, so the count is a close approximation and the tool marks it as an estimate. For exact numbers, Anthropic's API includes a token counting endpoint. ### GPT-5 token counter GPT-5 uses OpenAI's o200k_base encoding, and this tool runs that exact encoding in your browser. The counts match what OpenAI's own tiktoken library produces, for GPT-5 as well as GPT-4o and GPT-4.1. ### Gemini token counter Google does not publish Gemini's tokenizer. The tool uses the tokenizer of Gemma 3, Google's open model family, as a close proxy, and marks the result as an estimate. For exact numbers, the Gemini API offers a countTokens method. ## Frequently asked questions ### How many tokens is a word? **In English prose, one word is about 1.1 to 1.2 tokens for every model family we measured.** Our fixed 500-word sample produced 571 to 581 tokens across eleven tokenizers. Short common words usually cost one token; longer or rarer words split into more pieces, so technical writing and other languages push the ratio up. ### Is a token 4 characters? **Roughly, for English: OpenAI's published rule of thumb is that one token is about 4 characters.** On our plain 500-word sample the measured figure came out slightly higher, between 4.8 and 4.9 characters per token depending on the family. Treat 4 characters as a rough planning figure and measure your actual text when the number matters. ### Is there a tokenizer for Claude? **Anthropic has not published the tokenizer for current Claude models, so no public tool can give exact Claude counts.** This counter uses a close approximation and labels it as an estimate. When you need exact numbers, Anthropic's API provides a token counting endpoint that returns the count without running the model. ### How do I count tokens for GPT-5? **Paste your text into the counter with GPT-5 selected: it runs o200k_base, the same encoding GPT-5 uses.** Nothing is uploaded; the tokenizer runs in your browser. In code, OpenAI's open source tiktoken library gives the same counts with the o200k_base encoding.