About 1,333 tokens is a useful estimate for 1,000 words of English, using OpenAI’s rough guideline that 100 tokens correspond to about 75 words. It is only a planning estimate: the exact count depends on the model, tokenizer, language, and request format.
How many tokens are in common word counts?
The table applies the 0.75-words-per-token rule of thumb to English text. These are arithmetic estimates, not separate measurements or guaranteed counts. For example, 1,000 ÷ 0.75 gives about 1,333 tokens. OpenAI describes the relationship as approximately 100 tokens to 75 words in its guide to understanding and counting tokens.
As an Amazon Associate I earn from qualifying purchases.
| English word count | Approximate tokens |
|---|---|
| 100 | 133 |
| 250 | 333 |
| 500 | 667 |
| 750 | 1,000 |
| 1,000 | 1,333 |
| 1,500 | 2,000 |
| 2,000 | 2,667 |
Why the actual token count changes
Model and tokenizer
Tokens are pieces of text, not a fixed number of words. Depending on the tokenizer, a token may be a character, part of a word, a whole word, or punctuation. The same text can therefore produce different counts under different models or encodings. OpenAI recommends using the encoding associated with the model you plan to use when tokenizing programmatically; its Tokenizer can show how text is divided.
Language, wording, and formatting
The English estimate does not necessarily carry over to other languages. Spelling, capitalization, spaces, and punctuation can affect how a string is divided into tokens. A word counter and a tokenizer are measuring different things, so a word count cannot establish an exact token count.
#1 Best Overall
Provider-specific changes
Tokenizer behavior can change across model generations, even within one provider. Anthropic says its newer tokenizer for Claude 4.7 and later models and Claude Mythos Preview produces approximately 30% more tokens for the same text than earlier Claude models; the exact change depends on content and workload. Check the count against the specific model you intend to use rather than carrying over a count from an older model.
Everything in the request, not just visible prose
A prompt sent to an API may include message roles and boundaries, tool definitions, schemas, images, files, or prior conversation history. Those elements can add tokens beyond the text you see in a draft. OpenAI’s token-counting documentation explains counting supported structured inputs, including message framing; Anthropic’s token-counting documentation describes counting structured message inputs using the specified Claude model.
Quick Recap
How to count tokens for an LLM prompt
- Use the word-count estimate only for rough planning. For 1,000 English words, start with approximately 1,333 tokens, then verify if the limit matters.
- For plain text, use the target model’s tokenizer. OpenAI provides its Tokenizer and recommends
tiktokenfor programmatic plain-text counts. Select the encoding that matches the model rather than treating one tokenizer’s result as universal. - For an API request, count the structured input where supported. OpenAI’s Responses input-token counting endpoint accepts supported request inputs and accounts for structural tokens such as message roles and boundaries. Anthropic’s token-counting endpoint uses the specified Claude model tokenizer and accepts structured messages. Consult the relevant provider documentation for supported input types and endpoint details.
- Check the model’s context and output limits. The prompt is only part of the available context budget; the generated response also needs room. Reasoning models may use output tokens that do not appear in the final visible answer.
- Use reported usage to confirm the submitted call. After sending a request, check the provider’s usage fields for the actual input and output token counts. That result applies to that call, not automatically to differently formatted text or another model.
What to use when planning a prompt
- For a quick estimate of English prose, use roughly 1.33 tokens per word.
- For a tighter estimate, tokenize with the model and encoding you will use.
- For structured requests, count the complete supported request rather than only its prose.
- When staying under a context limit, budget for the answer as well as the input.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Recommended Free Tools
