How Many Words Is 1 Million Tokens?

Tokens to Words: The Conversion Table

Start here if you just need the number. Every row uses 0.75 words per token for English prose, 250 words per double-spaced page, and the 238 words-per-minute adult silent reading average from the Brysbaert 2019 meta-analysis of 190 studies.

For an exact count on your own text rather than an estimate, paste it into the AI prompt word counter, which shows the token count and per-model cost side by side, or use the tokens to words converter for the raw conversion in either direction.

What Is a Token, Exactly?

A token is the unit an AI model actually reads. It is not a word and it is not a character. It is a chunk of text, usually three to four characters, produced by a tokenizer that splits your writing into pieces the model has seen often enough to recognize.

Common short words are one token each. "The", "and", "is" all cost a single token. Longer or rarer words get split. "Tokenization" might become "token" plus "ization", two tokens for one word. Punctuation and spaces count too, and a leading space is usually bundled into the token that follows it, which is why "hello" and " hello" can tokenize differently.

This matters because every limit you deal with is measured in tokens, never in words. Context windows, output caps, rate limits, and API bills are all token-denominated. Your draft is measured in words. The gap between those two units is where people blow past a limit they thought they were under, and it is the entire reason this conversion is worth memorizing.

How Do You Convert 1 Million Tokens Into Words?

Multiply tokens by 0.75. One million tokens times 0.75 gives 750,000 words. Going the other way, multiply words by 1.33 to get tokens, so a 2,000-word article costs roughly 2,660 tokens.

The 0.75 figure comes from OpenAI's own documentation, which puts one token at about four characters of common English and 100 tokens at about 75 words. Since average English words including the trailing space run close to 5.3 characters, the arithmetic lands right where you would expect: four characters per token divided by 5.3 characters per word gives about 0.75.

Two shortcuts are worth keeping in your head. Tokens are roughly one third more than words, and words are roughly three quarters of tokens. Neither is exact, and neither needs to be. If you are within five percent of the real count, you will make the same decision about which model and which window to use.

Does the Ratio Change Between Models?

Yes, and this is the part most conversion charts skip. Every lab trains its own tokenizer, so the same paragraph produces a different token count depending on who is counting. OpenAI's o200k tokenizer is the most efficient of the major ones on English prose. Anthropic and Google tokenizers typically produce 10 to 20 percent more tokens for identical text.

The practical read: use 0.75 words per token when you are estimating, and use the real tokenizer when money or a hard limit is on the line. A 15 percent tokenizer difference on a 1 million token job is 150,000 tokens, which is a real bill and, more painfully, the difference between a document fitting and getting truncated.

How Big Is Every AI Context Window in Words?

Here is the 2026 lineup with each advertised context window converted into English words at the generic 0.75 ratio. The context window is the total budget shared by your input and the model's output, so a 200,000-token window does not give you 200,000 tokens of room for the prompt alone.

Per-model detail, including output caps and message limits, lives on the individual pages: Claude word limit, ChatGPT word limit, and Gemini word limit. The full set is indexed on the AI writing tools hub.

Can a Model Actually Use a Full 1M-Token Window?

Mostly no, and the gap is large enough to change how you plan a job. NVIDIA's RULER benchmark, which tests retrieval and reasoning at increasing input lengths, finds that effective context is typically only 50 to 65 percent of the advertised size. Accuracy degrades well before the ceiling.

The MRCR v2 multi-round retrieval test tells the same story with sharper numbers. At a full 1 million tokens the strongest frontier model holds roughly 76 percent accuracy. Several others drop below 50 percent, and Llama 4 Scout's advertised 10 million token window scores in the mid teens once the input actually gets that long. Meta itself acknowledges recall degrades far below the ceiling.

So treat the advertised window as a hard ceiling rather than a working capacity. If a model claims 1 million tokens, plan real work around 500,000 to 650,000, which is roughly 375,000 to 490,000 words. Put the material you actually need the model to use near the start and the end of the prompt, since the middle is where retrieval fails first. And when a document genuinely will not fit, chunking with retrieval still beats stuffing the window, no matter how big the number on the spec sheet is.

Does Content Type Change the Token Count?

Heavily. The 0.75 ratio holds for ordinary English prose and falls apart everywhere else. Anything unusual to the tokenizer costs more tokens per word.

Code is the trap that catches most developers. Indentation, brackets, camelCase identifiers, and symbols all fragment into separate tokens, so a 5,000-word codebase can cost 9,000 tokens instead of the 6,650 you budgeted. Markup is worse still. If you are pasting a scraped HTML page into a model, strip the tags first and you will often cut the token count by half.

Non-English text carries a real penalty too, since these tokenizers were trained mostly on English. Chinese and Japanese frequently tokenize close to one token per character, which is why the same document costs noticeably more to process in those languages. If you write in Spanish, our contador de palabras counts words natively before you convert.

What Does 1 Million Tokens Cost?

API pricing is quoted per million tokens, so this question has an unusually clean answer: 1 million input tokens costs exactly the listed input rate. Output is where the bill grows, since every major provider charges several times more for generated tokens than for the ones you send.

Two costs catch people out. First, reasoning or "thinking" tokens bill as output on models that use them, which means a short visible answer can carry a long invisible bill. Second, in a multi-turn conversation the entire history is resent with every message, so a chat that reaches 100,000 tokens of context charges you for 100,000 input tokens on each new turn, not just once. The AI prompt word counter estimates the per-model cost of a prompt before you send it.

How Many Words Is 100k Tokens, 128k, or 200k?

These three come up constantly because they are the windows most people actually work in. 100,000 tokens is about 75,000 words. 128,000 tokens is about 96,000 words. 200,000 tokens is about 150,000 words. Put a familiar object next to each and the scale stops being abstract.

One useful anchor: an average published novel runs about 90,000 words, which is roughly 120,000 tokens. So any model advertising a 128,000-token window can hold a whole novel, and a 1 million token window can hold a shelf of them. Whether it can reason across that shelf is a separate question, answered above. For book-length word counts by title, see our word count of popular books reference.

How Do You Count Tokens Before You Send Them?

Estimate first, verify when it matters. For a rough number, count your words and multiply by 1.33. For anything with a hard limit or a real bill attached, run the text through a tokenizer that matches your model.

The fastest path for most writing work: paste the draft into the word counter for an exact word and character count, then convert. If you want both numbers at once, the AI prompt word counter shows tokens and estimated cost across the current model lineup as you type. For batch conversions in either direction, the tokens to words converter handles it.

Three habits save the most trouble. Leave headroom of 10 to 15 percent below any hard limit, because your estimate is an estimate and the model's reply eats from the same budget. Strip markup and boilerplate before pasting, since tags and repeated navigation text are pure token cost with no information value. And when a long conversation starts feeling expensive or forgetful, start a fresh one with a summary instead of dragging the whole history forward.

Why Tokens and Not Words?

It is a fair question, since every user thinks in words and every model bills in tokens. The reason is that tokens are what the model computes on. A transformer processes a sequence of numeric IDs, and the tokenizer is the thing that turns text into those IDs. Words are a human convention that varies by language, and plenty of languages do not put spaces between them at all.

Subword tokenization is the compromise. Splitting on characters would make sequences far too long. Splitting on whole words would produce a vocabulary too large to train and would choke on anything unseen. Chunks of three or four characters handle both problems: common words stay whole, rare words decompose into recognizable pieces, and a misspelling or a brand-new product name still resolves into something the model can work with.

Practically, this means the unit mismatch is not going away. Learn the conversion once, keep the 1.33 multiplier handy, and check the real tokenizer when the stakes justify it.

How Many Tokens Is a Typical Document?

Abstract numbers are hard to plan with. These are the documents people actually paste into a model, with their word counts and the token cost that follows from them.

The pattern worth noticing is how much room a 1 million token window really has for ordinary work. You could paste 500 blog posts into it. What fills a window fast is not documents, it is conversation history and pasted code. A single long debugging session with repeated file dumps can burn 200,000 tokens without producing a single document. Check any draft's exact length in the word counter, or see typical targets in our dissertation length guide.

Is a Bigger Context Window Always Better?

No. Bigger windows are a capability, not a strategy, and using one badly is slower, costlier, and less accurate than using a smaller one well.

Three costs come with filling a window. Latency rises, because the model has to attend across everything you sent. The bill rises linearly with input, and in a multi-turn chat it rises on every single turn. Accuracy falls in the middle of long inputs, which is the failure the RULER and MRCR results above describe. Stuffing 800,000 words into a prompt to answer a question about one paragraph is the most expensive way to get a worse answer.

Large windows earn their keep on genuinely whole-document tasks: reviewing an entire contract for internal contradictions, tracing a change across a codebase, summarizing a book with the structure intact. For lookup-style questions over a large corpus, retrieval still wins. Fetch the relevant 4,000 words and ask about those. You get a faster answer, a cheaper answer, and usually a more accurate one.

Frequently Asked Questions

Related Tools

Related Articles