1. Count with a proxy tokenizer
Your text is tokenized locally with the o200k_base encoding, a modern OpenAI-compatible tokenizer, via gpt-tokenizer in your browser.
Claude Token Calculator
Plan Claude prompts in your browser. Anthropic does not offer a local tokenizer, so this calculator estimates Claude tokens with a disclosed proxy method and labels every count Estimated — then it estimates API cost and checks context-window fit. Honest about estimation, by design.
Private by default · No signup · Local token counting where supported · Exact and estimated results clearly labeled
Your text is processed in your browser for supported local tokenizers. We do not store pasted prompts.
Model comparison
Compare the current prompt, expected output, and monthly call volume across supported Anthropic models.
| Model | Accuracy | Prompt tokens | Per-call cost | Monthly cost | Context fit |
|---|---|---|---|---|---|
| OpenAI · GPT-5.5gpt-5.5 | Exact | — | — | — | — |
| OpenAI · GPT-5.4gpt-5.4 | Exact | — | — | — | — |
| OpenAI · GPT-5.4 minigpt-5.4-mini | Exact | — | — | — | — |
| OpenAI · GPT-5.4 nanogpt-5.4-nano | Exact | — | — | — | — |
| Anthropic · Claude Fable 5claude-fable-5 | Estimated | — | — | — | — |
| Anthropic · Claude Opus 4.8claude-opus-4-8 | Estimated | — | — | — | — |
| Anthropic · Claude Sonnet 5claude-sonnet-5 | Estimated | — | — | — | — |
| Anthropic · Claude Sonnet 4.6claude-sonnet-4-6 | Estimated | — | — | — | — |
| Anthropic · Claude Opus 4.6claude-opus-4-6 | Estimated | — | — | — | — |
| Anthropic · Claude Haiku 4.5claude-haiku-4-5-20251001 | Estimated | — | — | — | — |
| Google · Gemini 3.5 Flashgemini-3.5-flash | Estimated | — | — | — | — |
| Google · Gemini 3.1 Pro Previewgemini-3.1-pro-preview | Estimated | — | — | — | — |
| Google · Gemini 3.1 Flash-Litegemini-3.1-flash-lite | Estimated | — | — | — | — |
| Google · Gemini 3 Flash Previewgemini-3-flash-preview | Estimated | — | — | — | — |
| Google · Gemini 2.5 Progemini-2.5-pro | Estimated | — | — | — | — |
| Google · Gemini 2.5 Flashgemini-2.5-flash | Estimated | — | — | — | — |
| Google · Gemini 2.5 Flash-Litegemini-2.5-flash-lite | Estimated | — | — | — | — |
| Mistral · Mistral Largemistral-large-latest | Estimated | — | — | — | — |
| Mistral · Mistral Mediummistral-medium-latest | Estimated | — | — | — | — |
| Mistral · Mistral Smallmistral-small-latest | Estimated | — | — | — | — |
| Mistral · Ministral 14Bministral-14b-latest | Estimated | — | — | — | — |
| DeepSeek · DeepSeek V4 Flashdeepseek-v4-flash | Estimated | — | — | — | — |
| DeepSeek · DeepSeek V4 Prodeepseek-v4-pro | Estimated | — | — | — | — |
| xAI · Grok 4.3grok-4.3 | Estimated | — | — | — | — |
| xAI · Grok Build 0.1grok-build-0.1 | Estimated | — | — | — | — |
| Meta · Llama 4 Scoutmeta-llama/Llama-4-Scout-17B-16E-Instruct | Estimated | — | — | — | — |
| Meta · Llama 4 Maverickmeta-llama/Llama-4-Maverick-17B-128E-Instruct | Estimated | — | — | — | — |
Estimated rows use one local OpenAI-compatible proxy token count, so real provider counts can differ. Pricing estimate based on public model pricing. Check provider pricing before production use.
Why Claude counts are Estimated — and why that's honest
Only OpenAI encodings can be counted exactly in the browser today. Anthropic's official token counting is API-only, so this tool estimates Claude tokens with an OpenAI-compatible proxy and labels the result Estimated. We would rather show a clearly-labeled estimate than a falsely "exact" number.
2. Label it Estimated
Anthropic does not publish a local tokenizer, and Claude's tokenizer differs from OpenAI's, so the count is clearly labeled Estimated — never Exact.
3. Expect roughly ±10–20%
For most text the proxy lands within about ±10–20% of Claude's real count. Use it to plan context fit and cost, then confirm against Anthropic before production.
4. Stay private
Anthropic's exact counts come from its API, which would mean sending your text to a server. This privacy-first tool estimates locally instead, so your prompt stays in your browser.
The method, in one sentence
Estimated using an OpenAI-compatible tokenizer as a proxy. Anthropic's own tokenizer differs, so real counts typically vary by roughly ±10–20%.
Claude token calculator FAQ
Answers about Estimated Claude counting, the proxy method, and cost planning.
Is this calculator exact?
It is Exact only for supported OpenAI encodings counted locally with gpt-tokenizer. Other providers are labeled Estimated and use an OpenAI-compatible tokenizer as a proxy, so real provider counts can vary.
Why do different models show different token counts?
Different model families use different tokenizers. The same prompt can split into different token pieces for OpenAI, Claude, Gemini, Llama, Mistral, DeepSeek, or Grok models.
What is an LLM token?
A token is a chunk of text a language model reads or writes. A token can be a whole word, part of a word, punctuation, whitespace, a number, or part of code.
What is a context window, and what happens if I exceed it?
A context window is the maximum number of tokens a model can consider across input and expected output. If your total exceeds that limit, you need to shorten the prompt, reduce output length, or choose a larger-context model.
What is cached input pricing?
Some providers discount repeated input that can be reused from cache. When a model publishes cached-input pricing, the calculator shows it separately. If a model does not publish it, the row stays empty or unavailable.
Can I estimate monthly API cost?
Yes. Enter expected output tokens and calls per month. The calculator combines those values with the selected model's public pricing to estimate per-call and monthly cost.
Is my prompt uploaded to your server?
No. Pasted text is processed in your browser. This site has no accounts, saved prompt history, analytics scripts, ad scripts, or server-side text processing.
More token tools
Compare every provider on the main LLM token calculator, or see the Exact story on the OpenAI token calculator.