How many words is 1,000 tokens?

About 750 English words, using the same 0.75-words-per-token rule of thumb in reverse. Think of 1,000 tokens as roughly a page and a half of a typical paperback.

Is 1 token always 4 characters?

No — 4 characters per token is an average for English prose, not a law. Common words can be a single token (" the" is one token in many models), while rare words split into pieces and code or non-Latin scripts use far more tokens per character.

Do all AI models count tokens the same way?

No. Each model family ships its own tokenizer with its own vocabulary: OpenAI models use tiktoken-based BPE tokenizers, Claude and Llama families use different ones. The same text produces different token counts on different models, so always estimate against the model you'll actually call.

Why does code use more tokens than prose?

Tokenizers are trained mostly on natural language, so code's punctuation, indentation, and unusual identifiers split into many small pieces. A line of code that reads as 10 "words" to you might be 20+ tokens. Minified or dense code is the worst case.

What is a context window?

The maximum number of tokens a model can consider in one request — input plus the output it generates. A 128K-token window holds roughly 96,000 English words. Everything counts: your system prompt, conversation history, and the model's own reply.

How do I reduce token usage?

Trim conversation history (summarize instead of resending full transcripts), drop redundant instructions, avoid pasting the same context repeatedly, and prefer concise formatting over verbose scaffolding. Measure with an estimator before and after so you know what actually helped.

How Many Tokens Is 1,000 Words? LLM Token Counts Explained

September 26, 2026 · AI & LLMs

If you work with large language models, you'll keep bumping into the same question: how many tokens is my text? API pricing, context-window limits, and rate limits are all denominated in tokens — not words, not characters. Here's the short answer, the honest caveats, and a conversion table you can actually use.

The short answer: 1,000 words ≈ 1,300 tokens

The industry's standard rule of thumb for English prose:

1 token ≈ 0.75 words (about 4 characters)
1,000 words ≈ 1,300 tokens
1,000 tokens ≈ 750 words — roughly a page and a half of a paperback

So a 3,000-word blog post is ~4,000 tokens, and a 100,000-token context window holds around 75,000 English words — a short novel. These ratios are good enough for budgeting and "will it fit?" checks. They are not exact counts, and the rest of this article is about when they break.

What a token actually is

Models don't read words or characters — they read tokens, which are chunks of text from a fixed vocabulary learned during training. Common words get their own token; rare words are split into pieces. The word unbelievable might become three tokens — un, believ, able — while the is a single token including its leading space.

This subword approach (usually byte-pair encoding, BPE) is a compression trick: frequent strings get short codes, rare strings get spelled out in pieces. It's why token counts feel unpredictable — the tokenizer is optimized for the statistical patterns of its training data, not for human intuition about words.

The conversion cheat sheet

Words	≈ Tokens	Tokens	≈ Words
100	130	1,000	750
500	670	4,000	3,000
1,000	1,300	8,000	6,000
5,000	6,700	32,000	24,000
10,000	13,000	128,000	96,000

Read the left side when estimating input size ("my 5,000-word document is ~6,700 tokens"); read the right side when interpreting model limits ("128K context ≈ 96,000 words"). Remember: these are English-prose estimates. The next section is where they stop working.

Estimate your own text: paste anything into our free AI Token Estimator for an instant token count with GPT-style and Claude-style estimates, plus optional cost math.

Where the ratio breaks

The 0.75 rule assumes ordinary English prose. It degrades in predictable ways:

Code: tokenizers trained on prose shred code's punctuation and indentation into many small tokens. Dense code can run 2–3× the prose ratio — sometimes worse for minified JavaScript.
Non-Latin scripts: languages like Chinese, Arabic, or Hindi often tokenize at far more tokens per word than English — sometimes dramatically so — because the tokenizer's vocabulary is English-heavy.
Numbers and tables: long digit strings and tabular data split awkwardly; a spreadsheet pasted as text is token-expensive relative to its information.
Emoji and special characters: many emoji become multiple tokens each.

Practical takeaway: if your prompt is mostly code or non-English text, multiply the prose estimate by 2 as a safety margin — or better, measure the real thing.

Different models count differently

There is no universal token. Each model family trains its own tokenizer with its own vocabulary and merge rules: OpenAI's models use tiktoken-based encoders, Anthropic's Claude uses its own, Meta's Llama family uses SentencePiece-derived ones. The same paragraph yields different token counts on different models — usually within ~10–20% of each other for English, but diverging more on code and other languages.

This matters twice: once for cost (you're billed in the provider's tokens), and once for context limits (a prompt that fits in one model's window might not fit in another's). Always estimate against the model you'll actually call.

Why tokens matter: context windows and API costs

Two budgets, both in tokens:

Context window — the maximum tokens per request, covering everything: system prompt, conversation history, your new input, and the model's reply. Exceed it and the request fails (or gets truncated). Long documents and chat histories are the usual culprits.
API cost — billed per token, with output tokens typically priced higher than input tokens. A "cheap" prompt that generates a long answer can cost more than a long prompt with a short answer.

The most common budgeting mistake is counting only your input. A chatbot that resends full conversation history every turn pays for that history on every turn — which is why summarization and sliding windows exist.

How to get an exact count

Rules of thumb are for planning; exact counts are for shipping. Three options:

Use an estimator (like ours) for a fast approximation with clearly labeled assumptions.
Run the real tokenizer — OpenAI publishes tiktoken; other providers document theirs. This gives exact counts for that model family.
Ask the API — most providers return usage (prompt tokens, completion tokens) in every response. Log it; it's the ground truth for your bill.

For day-to-day work, the workflow is: estimate with the rule of thumb, verify with a real tokenizer before anything expensive, and reconcile against API-reported usage afterward.

Frequently asked questions

How many tokens is 1,000 words?: Roughly 1,300 tokens for ordinary English prose. The widely used rule of thumb is 1 token ≈ 0.75 words, so divide words by 0.75 (or multiply by 1.33). It's an estimate, not a guarantee — code and non-English text tokenize less efficiently.
How many words is 1,000 tokens?: About 750 English words, using the same 0.75-words-per-token rule of thumb in reverse. Think of 1,000 tokens as roughly a page and a half of a typical paperback.
Is 1 token always 4 characters?: No — 4 characters per token is an average for English prose, not a law. Common words can be a single token (" the" is one token in many models), while rare words split into pieces and code or non-Latin scripts use far more tokens per character.
Do all AI models count tokens the same way?: No. Each model family ships its own tokenizer with its own vocabulary: OpenAI models use tiktoken-based BPE tokenizers, Claude and Llama families use different ones. The same text produces different token counts on different models, so always estimate against the model you'll actually call.
Why does code use more tokens than prose?: Tokenizers are trained mostly on natural language, so code's punctuation, indentation, and unusual identifiers split into many small pieces. A line of code that reads as 10 "words" to you might be 20+ tokens. Minified or dense code is the worst case.
What is a context window?: The maximum number of tokens a model can consider in one request — input plus the output it generates. A 128K-token window holds roughly 96,000 English words. Everything counts: your system prompt, conversation history, and the model's own reply.
How do I reduce token usage?: Trim conversation history (summarize instead of resending full transcripts), drop redundant instructions, avoid pasting the same context repeatedly, and prefer concise formatting over verbose scaffolding. Measure with an estimator before and after so you know what actually helped.