Features
- Live token count for 100+ models across 12 providers (OpenAI, Anthropic, Google, Meta, xAI, Mistral, DeepSeek, Cohere, Alibaba, Moonshot, Perplexity, MiniMax)
- Side-by-side comparison so you can see how the same prompt tokenizes differently per model
- Cost estimate using each provider’s published per-million input/output rates
- Context-window meter shows how much of the model’s limit your prompt uses
- Handles code, JSON, markdown and non-Latin scripts (CJK, Arabic, Cyrillic) accurately
- Per-model calibration tuned against each provider’s published tokenizer where available
- Counts update as you type — no submit button, no waiting
- 100% client-side: prompts never leave your browser, no logging, no analytics on input
- Dark mode and offline PWA — works without internet after first load
How it works
- Open freetokencounter.app in any browser
- Paste or type your prompt into the input box
- Pick the AI model you’re targeting (or compare several side by side)
- Watch the live token count, estimated cost and context-window meter update
- Edit until you’re under the model’s context limit and within budget
Common use cases
- Estimating API costs before sending a long prompt to GPT-5 or Claude Opus 4.7
- Trimming a prompt to fit Gemini 2.5 Pro’s or Claude’s 1M context window
- Comparing tokenizers across providers when porting prompts between models
- Sizing system prompts and few-shot examples for a Llama 4 or Grok 4 deployment
- Quick-checking whether a code blob, JSON payload or transcript fits a context window
How it compares
OpenAI’s tokenizer page only counts tokens for OpenAI models, and most other counters are stuck on cl100k or GPT-3.5 estimates. freetokencounter.app supports 100+ models across 12 providers with per-model calibration, side-by-side comparison, live cost estimates and a context-window meter — all client-side. No sign-up, no upload, no API keys.
Privacy
Your prompt never leaves the browser. freetokencounter.app makes zero network calls after the page loads, runs no analytics on input, and stores nothing server-side. Safe for proprietary prompts, customer data and unreleased product copy — you can verify in the Network tab.
Frequently asked questions
What is a token in an AI model?
A token is the smallest unit of text an AI model processes. Tokens can be whole words, sub-words, or single characters depending on the tokenizer. As a rough rule, 1 token ≈ 4 characters or ¾ of an English word, but the exact count depends on the specific model. freetokencounter.app shows you the count for any major model in real time.
How accurate is freetokencounter.app?
Counts are estimates within roughly ±5% for English text on most models. The tool uses the public GPT-2 / cl100k splitting pattern combined with per-model calibration constants tuned against each provider’s published tokenizer. Some providers (Anthropic, Google, MiniMax) do not publish a fully open tokenizer, so counts for those models are clearly labeled as approximations.
Which models does freetokencounter.app support?
Over 100 models across 12 providers: OpenAI (GPT-5 family, GPT-4.1, o4-mini, o3, GPT-4o), Anthropic (Claude Opus 4.7, Sonnet 4.6/4.5, Haiku 4.5), Google (Gemini 3.1 Pro/Flash-Lite, 3 Flash, 2.5 Pro/Flash), Meta (Llama 4 Maverick/Scout/Behemoth, Llama 3.x), xAI (Grok 4, 4 Heavy, 4 Fast), Mistral (Medium 3, Small 3.1, Magistral, Pixtral, Codestral), DeepSeek (V3.1, R1), Cohere (Command A, R+, R), Alibaba (Qwen3-Max, Qwen3-Coder), Moonshot (Kimi K2), Perplexity (Sonar family), MiniMax (M1, Text-01).
Is my prompt uploaded anywhere?
No. freetokencounter.app runs entirely in your browser. Your prompt text never leaves your device — there are no servers, no analytics on input, no logging. You can verify this by checking the Network tab in your browser’s developer tools while typing.
Why does the same text produce different token counts on different models?
Each model family uses a different tokenizer trained on different data with a different vocabulary. GPT-5 and GPT-4o use o200k_base (~200K vocab), Claude Opus 4.7 uses Anthropic’s proprietary tokenizer, Gemini 2.5 Pro uses SentencePiece, Llama 4 uses its own BPE, and Grok 4 uses xAI’s tokenizer. The same word may be one token in one model and three tokens in another. freetokencounter.app shows the difference side by side.
How is cost calculated?
The cost estimate multiplies your input token count by each model’s published per-million input rate, plus an estimated output token count by the per-million output rate. Pricing data is sourced from each provider’s official pricing page and updated periodically. Always check live pricing for production budgeting.
What is a context window?
The context window is the maximum number of tokens (input + output combined) a model can process in a single request. GPT-5 handles 400,000 tokens, Claude Opus 4.7 up to 1 million, Gemini 2.5 Pro 1 million, Llama 4 Scout up to 10 million. If your prompt plus expected output exceeds the context window, the model will reject the request or truncate. freetokencounter.app shows a live context-window meter for the selected model.
Can I count tokens for code or non-English text?
Yes. freetokencounter.app handles code, JSON, markdown, and any Unicode text including non-Latin scripts. Note that code and non-English text typically use 30–80% more tokens than equivalent English prose, because their characters fall outside the most common subword merges. The tool shows accurate counts for any text you paste.