Token to words calculator
Tokens and words are not the same thing, and the ratio is not one number — it depends on the language, the content and the model's vocabulary. This page converts both directions, measures your own text with a real tokenizer, and shows the ratio each model family actually uses for English.
These are estimates for English prose. Code, JSON, numbers and non-Latin scripts use more tokens per word — paste the real text below to measure.
Measure your own text
Runs the GPT-4o / GPT-5-compatible BPE in your browser — nothing is uploaded. Exact for OpenAI models, within roughly 10% for Claude, Gemini, Llama and Qwen.
Token and word counts appear here as you type.
What the ratio actually is
For ordinary English prose, one token is a bit more than half a word: the widely used rule of thumb is 1 token ≈ 0.75 words, or 1 word ≈ 1.33 tokens. That number comes from averaging over large text corpora. It is accurate enough for budgeting a page, and wrong enough to be dangerous when you are trying to squeeze a prompt under a hard context limit — there, measure.
The reason the ratio is not exact is that tokenizers split text into sub-word units. “unbelievable” is three or four tokens, while “the” is one. Frequent words get their own token; rare words, numbers, punctuation clusters and whitespace patterns are broken into pieces. That is why the ratio shifts with writing style:
| Content type | Tokens per word | Words per 10K tokens |
|---|---|---|
| Plain English prose | 1.30 – 1.35 | 7,400 – 7,700 |
| Technical / academic English | 1.40 – 1.55 | 6,450 – 7,150 |
| Python or TypeScript source | 1.60 – 2.20 | 4,550 – 6,250 |
| JSON / config files | 2.00 – 3.00 | 3,300 – 5,000 |
| Chinese / Japanese text | 0.6 – 1.2 per character | ≈ 8,000 characters ≈ 5,000+ words |
| Digits, IDs, hashes | 3.00 – 5.00 | 2,000 – 3,300 |
Ranges, not constants. A token count is only exact when measured with the target model's own tokenizer.
Ratio by model family (English)
OpenAI, Anthropic, Google and Meta each ship a different vocabulary, so the same paragraph does not tokenize to the same number everywhere. These are the ranges the popular tokenizers land in on English prose; treat the spread as the error bar of a cross-model estimate.
| Model family | Tokens / word | Words / token | Chars / token |
|---|---|---|---|
OpenAI (GPT-4o / GPT-5, o200k family) Best measured with the exact same BPE used by the API — the calculator above does that. | 1.30 – 1.35 | 0.74 – 0.77 | ~4.0 |
Claude (Sonnet / Opus / Haiku) Claude tokenizes English slightly finer than OpenAI; counts land within ~10% of a GPT count. | 1.30 – 1.40 | 0.71 – 0.77 | ~3.8 |
Gemini (2.5 Pro / Flash) SentencePiece-based; a GPT tokenizer overestimates Gemini tokens by roughly 5–10% on prose. | 1.25 – 1.35 | 0.74 – 0.80 | ~4.0 |
Llama 4 / Qwen / DeepSeek / Mistral Non-Latin scripts and code diverge much more; always measure real input. | 1.30 – 1.45 | 0.69 – 0.77 | ~3.6 – 4.0 |
What a word count costs
Words only matter because tokens are what you are billed for. Using the ratio above, 10,000 words of English is roughly 13,300 tokens. At the list prices verified 2026-10-20:
| Model | Input / 1M | 10K words as input | Context window |
|---|---|---|---|
| GPT-5 | $1.25 | $0.0166 | 400K ≈ 301K words |
| GPT-5 nano | $0.05 | $0.0007 | 400K ≈ 301K words |
| Claude Sonnet 4.5 | $3 | $0.0399 | 200K ≈ 150K words |
| Gemini 2.5 Flash | $0.3 | $0.0040 | 1049K ≈ 788K words |
How to convert without a calculator
Words → tokens: multiply by 1.33 for English prose, 1.5 for technical writing, 2 or more for code and JSON. Tokens → words: divide by the same number, or multiply by 0.75. Characters → tokens: divide by 4 for English; for Chinese or Japanese, expect roughly one token per character. These three shortcuts are what most teams put in a spreadsheet before they bother calling a tokenizer.
When the number has a budget attached — a context limit you must not exceed, a per-request cost ceiling, a prompt-caching threshold — stop estimating and measure. The calculator above runs the tokenizer locally, so a paste is exact for OpenAI models and close for everything else.
Frequently asked questions
How many words is 1,000 tokens?
Around 750 words of ordinary English prose. The conversion is not a fixed number: English text averages about 1.33 tokens per word, so 1,000 tokens lands between 700 and 800 words. Code, JSON, URLs, numbers and non-Latin text carry far more tokens per word — 1,000 tokens of minified JSON can be under 200 “words”.
How many tokens is 1,000 words?
Roughly 1,300 to 1,400 tokens for plain English with normal punctuation. Short words, spaces and punctuation are cheap; long technical terms get split into several sub-word pieces, which pushes the ratio toward 1.5 or higher. Measuring the actual text with a tokenizer takes a second and removes the guesswork.
Is the tokens-to-words ratio the same for Claude and GPT?
No, but it is close for English. Anthropic, OpenAI and Google all use different BPE vocabularies, so the same paragraph can be 1,290 tokens for GPT-5 and 1,340 for Claude — a difference of a few percent. The ratio only becomes a real problem in non-English languages and code, where vocabularies differ much more.
How many tokens is a page of text?
A double-spaced page of 500 words is about 650–700 tokens. A dense single-spaced page of 600–700 words is about 850–950 tokens. For documents, estimate 1.33 tokens per word and you will be within about 10%.
Why do my token counts differ from the ones the API reports?
Because the tokenizer must match the model exactly. The calculator on this page runs the GPT-4o/GPT-5-compatible BPE in your browser, so OpenAI counts match the API. For Claude, Gemini, Llama and Qwen the vocabulary is different and counts are typically within 10% — good enough for budgeting, not for exact billing reconciliation.
How do I convert tokens back to a cost figure?
Multiply token count by the per-1M-token price of the model you are using. For example, 50,000 input tokens on a model priced at $1.25 per 1M input tokens costs $0.0625. The cost calculator on the home page does this for every model in the pricing table at once, including the output side.
Related tools
- Token counter & cost calculator — paste a prompt, count tokens and get per-request, monthly and yearly spend
- Model pricing table — input, output and cached-input prices for every model
- Budget calculator — how many requests a monthly budget buys
- Cheapest models — ranked by blended price per request