SynovixLabs
Blog

How Many Tokens Per Word? Complete Reference

Updated June 2026 · 5 min read

If you are building a ChatGPT prompt, estimating API costs, or trying to stay within a model's context window, you need a reliable answer to one question: how many tokens per word does an AI actually use?

Quick answer: 1 English word ≈ 1.3–1.5 tokens. So 1,000 words ≈ 1,300–1,500 tokens.

That single ratio — roughly 1.3 to 1.5 tokens per word — works well for plain English prose. But the real number shifts depending on whether you are feeding in code, scientific jargon, JSON, or non-Latin languages. This guide gives you a full reference table plus the reasoning behind every figure, so you can make accurate estimates without guessing.

Want an exact count instead of an estimate?
Paste your text into the free AI Token Calculator and get a precise token count for GPT-4, Claude, and Gemini models instantly — no sign-up required.

Tokens per word: the core formula

AI language models do not read words directly. They first run your text through a tokenizer — an algorithm that breaks input into subword chunks called tokens. The result depends on the model's vocabulary and the specific text, but for English prose the pattern is consistent:

  • 1 token ≈ 4 characters of English text
  • 1 token ≈ ¾ of a word, or equivalently 1 word ≈ 1.33 tokens
  • 100 tokens ≈ 75 words — useful for context window math
  • Common short words ("the", "and", "of") usually map to exactly 1 token each
  • Longer or rarer words ("photosynthesis", "tokenization") often split into 2–3 tokens
  • Spaces, punctuation, and newlines each count as their own tokens
  • Numbers, especially large or decimal ones, can become multiple tokens
  • Code identifiers in snake_case or camelCase typically tokenize more aggressively than prose words

The ratio of 1.3–1.5 is a practical average across these cases. For a detailed explanation of why tokenization works this way, see What Are AI Tokens?

How many tokens is 1000 words?

Using the standard English prose ratio of 1.3–1.5 tokens per word, 1,000 words converts to approximately 1,300–1,500 tokens. A safe midpoint for planning is 1,400 tokens per 1,000 words.

To put that in context: GPT-4o has a 128,000-token context window. That means it can handle roughly 85,000–98,000 words in a single conversation — the length of a full novel. Claude 3.5 Sonnet supports 200,000 tokens, which is around 133,000–154,000 words.

Knowing this ratio also helps with API cost estimation. Most models price by tokens consumed (both input and output), so being able to convert your word count to a token count is the first step in forecasting your bill.

Words to tokens reference table

Use this table as a quick lookup for common word counts. The "min" column uses 1.3 tokens/word (typical short-word-heavy text) and "max" uses 1.5 tokens/word (technical or varied vocabulary).

Word count Min tokens Max tokens Example content
100 words 130 150 Short paragraph or tweet thread
200 words 260 300 Product description, short bio
500 words 650 750 News article intro, short blog post
1,000 words 1,300 1,500 Standard blog post, landing page
2,000 words 2,600 3,000 Long-form article, detailed tutorial
5,000 words 6,500 7,500 Short ebook chapter, white paper
10,000 words 13,000 15,000 Full ebook, research report
80,000 words 104,000 120,000 Full novel, fits GPT-4o context

Note: these are estimates for English prose. Code and non-English text will differ — see the table below.

Tokens per word by content type

The 1.3–1.5 ratio only holds for standard English prose. Different content types tokenize very differently because tokenizers are optimized around common English subwords. Here is what to expect:

Content type Tokens per word (approx.) Why it differs
English prose 1.3–1.5 Vocabulary is well-covered by tokenizer; short words dominate
Technical / scientific English 1.5–2.0 Long compound terms, abbreviations, and Greek letters split into extra tokens
Python code 1.2–1.6 Short keywords ("if", "for", "def") are compact; long identifiers split
JSON / structured data 2.0–3.0 Braces, quotes, colons, and newlines each consume tokens; keys repeat overhead
Chinese text 1.5–2.5 per character Each CJK character is roughly 1–2 tokens; "words" are 2–4 characters
Arabic / Hebrew 3–5 per word Rich morphology and right-to-left script are poorly covered by English-centric vocabulary

The practical takeaway: if you are sending mixed content — a user message in English followed by a code block and some JSON — you should estimate each section separately using the appropriate multiplier, then sum them. The overall ratio could easily land at 2.0 or higher even if the prose portion looks normal.

For JSON-heavy payloads in particular, consider compressing whitespace before sending to the API. Removing indentation and extra newlines from a JSON payload can cut its token count by 30–40% with no loss of information.

How many tokens is a typical ChatGPT conversation?

Token usage accumulates across a conversation because most APIs send the full conversation history with every new message. Here is a realistic example of how tokens pile up in a multi-turn ChatGPT session:

  • System prompt (200 words of instructions): ~280 tokens
  • Turn 1 — user message (50 words): ~70 tokens
  • Turn 1 — assistant reply (300 words): ~430 tokens
  • Turn 2 — user follows up (30 words, but full history resent): ~780 tokens input, plus ~50 tokens new message
  • Turn 2 — assistant reply (250 words): ~360 tokens
  • Total after 2 turns: roughly 2,000+ tokens consumed

This compounding effect is why long multi-turn sessions can exhaust a model's context window — or balloon your API bill — faster than you expect. Every new turn sends not just the latest message but the entire conversation history up to that point. A 10-turn conversation with moderately detailed replies can easily consume 15,000–25,000 tokens.

To manage this, many production applications implement context window trimming: they keep the system prompt and the most recent N turns in full, then summarize or discard older turns. See how to reduce AI API costs for concrete strategies.

If you want to track token usage in real time or stress-test a prompt before deploying it, the AI Token Calculator lets you paste any text and get a precise count across models.

Frequently asked questions

How many tokens is 1000 words?

Using the standard English prose ratio, 1,000 words is approximately 1,300–1,500 tokens. A practical midpoint for planning or cost estimation is 1,400 tokens. For technical text or content that mixes languages, use a higher multiplier of 1.5–2.0 to be safe.

How many tokens is 500 words?

500 words is roughly 650–750 tokens for typical English prose. That is well within the comfortable range for a single API call. If your 500-word input is all plain narrative, 700 tokens is a reliable single-number estimate.

How many tokens per word does ChatGPT use?

ChatGPT (and the underlying GPT-4 family) uses the tiktoken tokenizer with the cl100k_base vocabulary. For English text this produces approximately 1.3–1.5 tokens per word — the same ratio you see cited in OpenAI's own documentation. The exact count depends on word length and frequency, but the ratio is very consistent across normal conversational text.

Do tokens count spaces and punctuation?

Yes. Spaces, punctuation marks, newlines, and special characters all count toward your token total. In the tiktoken tokenizer, a leading space is often merged into the following word token (so " hello" is one token), but punctuation and symbols usually each get their own token. This is why dense formatting — JSON, markdown headers, code with many brackets — adds tokens faster than plain sentences.

How do I count tokens without installing anything?

The easiest method is to use an online tool. The AI Token Calculator on SynovixLabs runs entirely in your browser — no sign-up, no install — and shows you the exact token count for GPT-4, Claude, and Gemini side by side. For a detailed walkthrough of how to use it, see the AI Token Calculator guide.

Related reading