Paste or type any text into the box above. Nothing you enter is sent anywhere; it's counted entirely in your browser.
02
Pick your model
Switch between GPT-5/GPT-4o, GPT-4/GPT-3.5, Claude and Gemini. Each provider tokenizes text differently, so the same prompt can produce a different count on each.
03
Watch the buffer fill
The context buffer meter shows how much of that model's typical context window your prompt is using, so you can see how much room you have left before you hit a limit.
No provider, not OpenAI, Anthropic, or Google, publishes a full offline tokenizer for their current flagship models. Getting a truly exact count would mean sending your prompt to their API, so instead this tool uses each provider's own published character-to-token ratio, the same approximation method they recommend when API access isn't available.
Alongside the token count, you'll always see the character and word count too, useful for converting tokens to words when you're budgeting a prompt, or as a quick token cost calculator once you multiply by your provider's current per-token rate.
Anthropic's tokenizer changed with Opus 4.7 and later, current Claude models tokenize about 30% more densely than older ones, so the Claude estimate here is calibrated to today's models, not the outdated ~4 characters per token figure still floating around online. GPT-5 uses the same o200k_base encoding as GPT-4o, and Gemini's ratio comes straight from Google's own current documentation.
Open-weight models are covered too, Qwen 3.6, DeepSeek V4, Llama 4, and Kimi K2, since Chinese open-weight labs have genuinely overtaken Western models in total downloads this year, and most token counters still ignore them entirely. Their tokenizers are all large-vocabulary BPE encoders that land close to the same 4-characters-per-token range on English text, Alibaba states this directly for Qwen; the others are calculated the same way. Kimi's figure is the least certain of the eight, since Moonshot hasn't published tokenizer-specific benchmarks the way the others have.
Everything runs client-side. Your prompt is never uploaded, logged, or stored; refresh the page and it's gone. No downloads, no waiting.
A token is the small chunk of text an AI model actually reads, usually a few characters or part of a word, rather than a whole word. The word "tokenization" itself is a good example: most tokenizers split it into two pieces, "token" and "ization", rather than treating it as one unit. Tokens matter because they're what you're billed for and what counts against a model's context window, not raw character or word counts.
A token counter measures how many tokens a piece of text will use once an AI model processes it, converting your raw characters and words to tokens so you can see usage and cost before you send a prompt. Tokens aren't quite words or characters; they're the small chunks that models like GPT, Claude and Gemini actually read, and that you're billed for through their APIs.
Paste your prompt into the box above and pick a model. This calculator gives you an instant estimate using each provider's own published character-to-token ratio, calibrated to their current models, with no download and no wait.
It's a close estimate, not a guaranteed exact count. OpenAI's own documentation cites roughly 4 characters per token for English text under both the o200k_base encoding (GPT-5, GPT-4o) and cl100k_base (GPT-4, GPT-3.5), which is what this tool uses. For a billing-critical exact count, OpenAI's tiktoken library is the authoritative source.
Anthropic hasn't published a public offline tokenizer for any current Claude model. The only way to get an exact count is Anthropic's own token-counting API, which needs an API key and a server call, something this tool avoids entirely to keep your prompts private. Instead, this uses a character-ratio estimate calibrated to current Claude models specifically.
Yes. Starting with Claude Opus 4.7, Anthropic's newer models tokenize text about 30% more densely than older ones, meaning the same prompt now uses more tokens than it would have on Claude 3.x or Opus 4.6. A lot of token estimate tools still quote the old ratio, roughly one token per 4 characters, which understates usage on current models like Sonnet 5 and Opus 4.8. This calculator uses the updated ratio.
Google's own current documentation states that a Gemini token is equivalent to about 4 characters, or roughly 60 to 80 English words per 100 tokens, which is the ratio this tool uses.
GPT-3.5 and GPT-4 both use the cl100k_base encoding. GPT-4o and GPT-5 switched to a newer encoding called o200k_base, with a larger vocabulary that handles non-English text and code more efficiently. For plain English text the two encodings land on a similar character-per-token ratio, which is exactly why this tool lets you switch between them for whichever family you're planning around.
Yes. Qwen 3.6, DeepSeek V4, Llama 4 and Kimi K2 are all included alongside GPT, Claude and Gemini. Most token counters skip open-weight models entirely, even though Chinese labs in particular have overtaken Western labs in total open-weight downloads this year. All four use large-vocabulary BPE tokenizers that land close to the same 4-characters-per-token range as GPT on English text; Alibaba states this directly in Qwen's own documentation. Kimi's figure is the one estimate here with the least direct sourcing, since Moonshot hasn't published tokenizer-specific benchmarks the way the other labs have.
It depends on the model tier and whether you're paying for input or output tokens; Anthropic's API prices each separately, and output tokens generally cost several times more than input. This tool doesn't calculate live pricing, since rates change often and vary by model, but counting your tokens first is the first step: multiply the count by your model's current per-token rate from Anthropic's pricing page to get a real cost estimate.
Roughly 1,500 pages, using the common publishing convention of about 500 words per page and the standard ~0.75 words-per-token ratio most providers cite. By the same math, 200K tokens works out to around 300 pages, and 128K tokens to around 190 pages. These are rough conversions, not exact, since actual word and page density varies by formatting and content type.