AI token counter & prompt workspace
Exact OpenAI token counts, an honest range for every other model, your own token budget, and a structural read on the prompt itself. Everything runs in your browser.
Check · does it fit your budget?
CountBudgets are yours to set. We do not publish per-model prices or context limits, because those change often and a stale number is worse than none.
- Countcounted in your text
- Estimatemodelled, shown as a range
Tokens are the exact o200k_base count (GPT-4o, GPT-5), computed locally. Loading the tokenizer — showing a character-ratio fallback.
- cl100k tokens
- 0
- Sentences
- 0
- Lines
- 0
- Avg. sentence
- 0.0 w
Other models
EstimateModel · 3.5–4.5 characters per token, the span current Claude, Gemini, DeepSeek and Llama tokenizers fall into for English prose. Not a published figure from those providers.
What a token actually is
Language models do not read characters or words. They read tokens: chunks of text the tokenizer has learned to treat as single units. A common English word is usually one token, a rare one splits into two or three, and punctuation, whitespace and emoji all carry their own cost. That is why a 1,000-character prompt is never reliably 250 tokens, and why counting characters is a poor substitute for counting tokens.
Where the numbers come from
For OpenAI encodings this page runs the real tokenizer. o200k_base is what GPT-4o and GPT-5 use; cl100k_base is what GPT-4 Turbo and the embedding models use. Both are executed locally through js-tiktoken, so the number matches the API exactly. Everything else is presented as a range, because Anthropic, Google, DeepSeek and Meta do not publish tokenizers we can run in a browser. A character-ratio band is the honest answer there, and it is labelled as one.
Budgets instead of price tables
Most token counters bolt on a pricing table. We do not, because published API prices and context limits move several times a year and a table nobody re-verifies quietly becomes misinformation. What does not go stale is your own constraint: the budget for a system prompt, a retrieval chunk, or a per-request ceiling your application enforces. Set that number and the workspace tells you exactly how much of it this prompt spends.
Structure matters as much as size
The structure panel flags what experienced prompt writers check anyway: a missing role instruction, no examples, no stated output format, sentences long enough to blur the instruction, and duplicated lines that spend tokens without adding information. None of these are rules. They are prompts to look again at a prompt.
Frequently Asked Questions
Counting words or characters instead?
Open the text workspace →