How Many Tokens Are in 1,000 Words?
Short answer: about 1,333 tokens for ordinary English prose. The useful range is 1,250–1,400 for plain prose, rising to 1,400–1,600 for technical English, and 1,900–3,000 for code, JSON and HTML. Below: the conversion table by content type, what 1,000 words actually costs to send and generate on every tracked model, and a calculator for your own word counts.
The 1.33x rule, and when it breaks
One English word averages about 1.33 tokens, or one token averages about 0.75 words. That ratio holds for prose — and breaks badly everywhere else.
Tokens per 1,000 words, by content type
Estimates for modern sub-word tokenizers. For code, markup and structured data, "words" means whitespace-delimited units — a practical planning proxy, not a linguistic one.
| Content type | Tokens per 1,000 words* | Words per 1,000 tokens | Why it differs |
|---|---|---|---|
Plain English proseBaseline Articles, chat, support text | 1,250 – 1,400 | ~750 | Short, common words usually map to one token each. |
Business / technical English Reports, specs, documentation | 1,400 – 1,600 | ~650 | Longer words, jargon, numbers and abbreviations split into sub-words. |
Transcribed speech Call transcripts, dictation | 1,200 – 1,450 | ~760 | Fillers and short words tokenize cheaply; punctuation is sparse. |
Source code Python, JavaScript, TypeScript | 1,900 – 2,400 | ~470 | Indentation, operators, camelCase boundaries and symbols all split. |
JSON / structured data API payloads, configs | 2,300 – 2,700 | ~410 | Every key is repeated, and each quote, brace and colon is its own token. |
HTML markup Scraped pages, emails | 2,500 – 3,000 | ~370 | Tags, attributes and nesting add tokens that carry no content. |
CJK text Per 1,000 characters, not words | 950 – 1,100 | — | Roughly one token per character; word segmentation is not meaningful. |
* Ranges are estimates for planning. Exact counts depend on the specific tokenizer each provider uses, and the same text can meter several percent differently across OpenAI, Anthropic and Google. Measure real text with the Token Counter before committing to a budget.
1,000 words in, 1,000 words out — priced on every model
Fixed volume: 1,000 words of input (≈1,333 tokens) and 1,000 words of output (≈1,333 tokens) per request, at standard per-1M rates. The "per 100,000 requests" column is what this looks like at production scale.
| Model | Input cost | Output cost | Per request | Per 100,000 requests | Output share of bill |
|---|---|---|---|---|---|
GPT-5.6 Sol OpenAI · flagship (promo) | $0.0053 | $0.0267 | $0.0320 | $3,199.20 | 83.3% |
GPT-5.6 Terra OpenAI · mid | $0.0027 | $0.0160 | $0.0187 | $1,866.20 | 85.7% |
GPT-5.6 LunaLowest OpenAI · budget | $0.0003 | $0.0016 | $0.0019 | $186.62 | 85.7% |
Claude Opus 5 Anthropic · flagship | $0.0067 | $0.0333 | $0.0400 | $3,999.00 | 83.3% |
Claude Sonnet 5 Anthropic · mid | $0.0027 | $0.0133 | $0.0160 | $1,599.60 | 83.3% |
Claude Haiku 4.5 Anthropic · budget | $0.0013 | $0.0067 | $0.0080 | $799.80 | 83.3% |
Gemini 3.6 Flash Google · mid (promo rate) | $0.0010 | $0.0050 | $0.0060 | $599.85 | 83.3% |
Gemini 3.5 Flash-Lite Google · budget | $0.0004 | $0.0033 | $0.0037 | $373.24 | 89.3% |
Computed at 1,333 input and 1,333 output tokens per request using published per-1M rates (OpenAI verified 2026-08-23, Anthropic and Google 2026-08-09 to 2026-08-16). GPT-5.6 Sol's promotional rate runs through at least Nov 21, 2026; Gemini 3.6 Flash's through Dec 31, 2026. Cheapest-to-dearest spread is 21x — the same 1,000 words costs $186.62 or $3,999.00 per 100,000 requests depending on routing. Caching, batch and long-context rules excluded.
Convert your own word count and price it
Enter word counts and a tokens-per-word factor — use a preset, or enter your own if you have measured your real content. Every tracked model is priced on the result.
| Model | Input cost | Output cost | Per request | Per 100,000 requests |
|---|
Tokenizers split sub-words, not words.
Modern LLM tokenizers break text into sub-word units chosen from a fixed vocabulary. Common English words like "the", "project" or "budget" usually occupy a single token. Rare words, product names, numbers and code identifiers get split into two or more — "tokenization" might become three tokens, and a UUID can become a dozen.
That is why the same word count produces wildly different token counts by content type. In JSON, every key is repeated on every record and each {, }, " and : bills separately. In HTML, tags and attributes add tokens that carry no information at all — stripping markup to clean Markdown routinely removes 30–60% of the tokens in web-derived content.
Practical consequence: measure the content you actually send. A rule of thumb is fine for scoping a budget; it is not fine for a per-customer price quote. Paste real documents into the Token Counter and check the result against your provider's usage API.
Where the money actually goes on a 1,000-word job.
Two numbers from the table above matter more than the conversion ratio. First, output is 83–89% of the bill on a balanced 1,000-in / 1,000-out request — the generation side dominates because providers price output 5–8x above input. Capping output length is usually a bigger saving than trimming the prompt.
Second, model routing is a 21x lever. The same 1,000 words costs $186.62 per 100,000 requests on GPT-5.6 Luna and $3,999.00 on Claude Opus 5. If a classification or extraction job does not need flagship reasoning, routing it to a budget model is worth more than any prompt optimization.
Finally, content type is a lever most teams ignore: converting scraped HTML to clean Markdown before sending can cut input tokens by 30–60%, which on an input-heavy RAG workload is a direct cost reduction.
Words-to-tokens questions
How many tokens is 1,000 words?
About 1,333 tokens for ordinary English prose, using the standard rule of thumb of 0.75 words per token. The realistic range is 1,250–1,400 tokens for plain prose, rising to 1,400–1,600 for technical or business English. Code, JSON and HTML run two to three times higher.
How many words is 1,000 tokens?
About 750 English words. That is the inverse of the same rule of thumb: 1 token is roughly 4 characters, or 0.75 words, of ordinary English. For technical prose it is closer to 650 words; for source code, around 470 whitespace-delimited units.
Why do different content types give different token counts for the same word count?
Tokenizers split on sub-word units, not whole words. Short common English words often become a single token, while jargon, numbers, punctuation, quotes and markup tags are split into several. Every brace, quote and angle bracket in JSON or HTML is billed separately, which is why structured content costs roughly two to three times more per thousand words than plain prose.
How much does it cost to send 1,000 words to an LLM?
Sending 1,000 English words (about 1,333 input tokens) costs between $0.0003 and $0.0067 depending on the model — from GPT-5.6 Luna at $0.0003 to Claude Opus 5 at $0.0067. Generating 1,000 words back costs 5–8x more, between $0.0016 and $0.0333, because output tokens are billed at a much higher rate.
How many 1,000-word documents fit in a 1M-token context window?
About 750. A 1,000,000-token context window holds roughly 750,000 English words at 1,333 tokens per 1,000 words. In practice you must subtract your system prompt, tool definitions and the output you want back, so budget for fewer — and confirm with a context-window check before sending a large batch.
Where these numbers come from.
Conversion ratios are planning estimates derived from how modern sub-word tokenizers behave — they are not exact for any specific model. Exact counts require each provider's own tokenizer, and the same text can meter several percent differently across OpenAI, Anthropic and Google. Cost figures multiply the estimated token counts by published per-1M rates taken from official provider pricing pages and re-verified weekly; they do not apply caching, batch or tiered discounts.
Official sources: OpenAI model docs, Anthropic pricing, Google Gemini API pricing. Always confirm the invoice before making purchasing decisions.