LLM Token Counter
Estimate prompt tokens locally with clear accuracy labels and model-specific tokenizer disclosures.
Plain text only · estimated locally · no provider API
PROMPT TEXT
TOKEN ESTIMATE
ESTIMATED TOKENS
—
GPT-6.1 Sol · text heuristic
No verified local GPT-6.1 encoding is bundled. This is a text heuristic, not the provider tokenizer.
Paste a prompt and count when ready.
Heuristic: ASCII characters ÷ 4, other UTF-8 bytes ÷ 2, rounded up. The same fallback is used for these models; their actual tokenizers differ. It is not a guaranteed bound. Non-English text, code and punctuation can differ substantially. Message wrappers, tools, media and output tokens are excluded.
Enable JavaScript to count locally. No server submission is available.
A useful estimate, with honest limits.
Count plain text, source code, JSON or multilingual prompts without uploading them. Choose the intended model to see its tokenizer disclosure. The current catalog uses a local text heuristic for every model; none is labeled exact.
Count the complete request elsewhere when billing matters.
Provider usage includes message wrappers, conversation history, tool definitions and potentially media. This editor counts only the text supplied. It cannot predict hidden reasoning or a future response.
What the result means.
Why no token coloring?
Reliable boundaries require the actual tokenizer. Whitespace splitting would misrepresent tokenization, so no token-boundary visualization is shown.
What are the limits?
64 KiB UTF-8, explicit Count tokens, no expensive work on each keystroke. Model changes recalculate an existing result. Reset or leaving the page discards prompts and counts.
Can I estimate cost or context usage?
Use the related calculators and enter the count manually. Prompt text and derived counts are never transferred through URLs or storage. Prefer actual provider usage counts for budget decisions.
Which languages work?
Unicode, emoji, Hindi and Gujarati are accepted. Character counts use Unicode code points, words use whitespace boundaries and empty input has zero lines. Token estimates can vary widely by language and content.