How Many Tokens Is a Word? (2026 Cheat Sheet)
Short answer: one English word averages 1.3 tokens with OpenAI's o200k tokenizer (roughly 4 characters per token, or ~3 tokens per 4 words). For Chinese text the ratio flips: 1 character ≈ 0.6–1.2 tokens. Exactly, token counts depend on the model's tokenizer, which is why a word counter is not a token counter.
Tokens per word by model family (2026)
| Model family | Tokenizer | Tokens per word | Chars per token |
|---|---|---|---|
| GPT-5 / GPT-4o / o1 / o3 | o200k_base | ~1.3 | ~3.9 |
| GPT-4 / GPT-3.5 (legacy) | cl100k_base | ~1.3 | ~3.9 |
| text-davinci-003 | p50k_base | ~1.25 | ~4 |
| Claude (Opus/Sonnet/Haiku) | estimate | ~1.25 | ~4 |
| Gemini (Pro/Flash) | estimate | ~1.25 | ~4 |
| Llama (4 / 3) | estimate | ~1.25 | ~4 |
| DeepSeek V3/R1 (Chinese) | estimate | ~0.6 per char | — |
Why the ratio is never exact
Tokenizers split text with Byte-Pair-Encoding, not by words:
- Common words like
the,and,thatare usually a single token. - Long or rare words split into several tokens (
internationalization≈ 4 tokens). - Whitespace is typically merged into the following token — so indentation and line breaks matter.
- Code costs more than prose: symbols like
{,},=>tokenize densely. - Emoji and non-Latin scripts behave very differently (Chinese: ~0.6–1 token per character).
Quick word → token estimates
| English words | ≈ Tokens (o200k) |
|---|---|
| 100 words | ~133 tokens |
| 500 words | ~665 tokens |
| 1,000 words | ~1,330 tokens |
| 5,000 words | ~6,650 tokens |
| 10,000 words | ~13,300 tokens |
Rule of thumb that still works: tokens ≈ characters ÷ 4 for English, tokens ≈ characters × 0.6 for Chinese.
“Tokens vs words” — why it matters
API providers bill by tokens, and models have hard token limits per context window. A 1,000-word prompt is ~1,330 tokens of your context budget on GPT-5, Gemini 2.5 Pro (1M window) or Claude 4 (200K window). The same prompt in Chinese is ~800–1,200 tokens depending on the tokenizer. Estimating with words instead of tokens can be off by 30% — enough to reject a prompt that doesn't fit, or to double your invoice.
Don't guess — count with the real tokenizer
Our AI Token Calculator runs the actual tiktoken tokenizer (the same one OpenAI uses) right in your browser — no upload, no API key. Paste your prompt, article or dataset and get invoice-accurate token counts for GPT-5, GPT-4o, o1, o3 and every other OpenAI model, plus proven estimates for Claude, Gemini, Llama, DeepSeek, Qwen and 90+ more.