How Many Words Is 1M Tokens?
Short answer: about 750,000 English words, or 1,500 pages at 500 words each. But 1M tokens is where the unit stops behaving like a document and starts behaving like a budget — only 5 of the 8 models tracked here can accept it in one request, and on OpenAI it crosses the 272,000-token threshold that doubles the input rate. Below: the conversion by content type, which models can actually hold 1M, and why the headline price comparison inverts at this size.
The 0.75 rule, at a million
One token averages about 0.75 English words. At 1,000,000 tokens the arithmetic is simple — the caveats are not.
What 1M tokens gives you, by content type
Content type changes the answer more than any other factor. For code, markup and structured data, "words" means whitespace-delimited units — a planning proxy, not a linguistic one.
| Content type | Words from 1M tokens | 500-word pages | Why it shrinks |
|---|---|---|---|
Plain English proseBaseline Articles, chat, support text | 714,000 – 800,000 | 1,430 – 1,600 | Common short words usually map to one token each. |
Business / technical English Reports, specs, documentation | 625,000 – 714,000 | 1,250 – 1,428 | Jargon, numbers and long words split into sub-words. |
Transcribed speech Call transcripts, dictation | 690,000 – 833,000 | 1,380 – 1,666 | Short filler words tokenize cheaply; punctuation is sparse. |
Source code Python, JavaScript, TypeScript | 417,000 – 526,000 | 834 – 1,052 | Indentation, operators, camelCase and symbols all split. |
JSON / structured data API payloads, configs | 370,000 – 435,000 | 740 – 870 | Keys repeat per record; each quote, brace and colon bills. |
HTML markup Scraped pages, emails | 333,000 – 400,000 | 666 – 800 | Tags and attributes add tokens that carry no words at all. |
CJK text Characters, not words | 909,000 – 1,053,000 chars | — | Roughly one token per character. |
Ranges invert the tokens-per-1,000-words figures published on our 1,000-words conversion table. They are planning estimates, not billing-grade: the same text meters several percent differently across OpenAI, Anthropic and Google tokenizers. Measure real documents with the Token Counter before committing to a budget.
What 1M tokens looks like in practice
Abstract token counts are hard to reason about. These are the equivalents teams actually plan against.
| Unit | 1M tokens is about… | Planning note |
|---|---|---|
English words 0.75 per token | 750,000 | Seven and a half 100,000-word novels. |
500-word pages Standard manuscript page | 1,500 | A 1,500-page report — far larger than any single deliverable. |
Lines of source code At ~8–12 tokens per line | 83,000 – 125,000 | A large monorepo module set, not a single file. |
Support conversations At ~2,000 tokens each | 500 | Roughly two weeks of tickets for a small team. |
Plain-text file size At ~4 characters per token | ≈ 4 MB | PDFs and HTML inflate well beyond this once markup is included. |
Filesize is the least reliable of these: a 4 MB PDF can meter at 1.5M tokens or 400K depending on whether it is text-native or scanned. Always extract and count the text.
Can 1M tokens even fit in one call?
This is where 1M stops being a document size and becomes a routing decision. Context windows are shared by input and output, so a window that exactly equals your prompt leaves nothing for the answer.
| Model | Context window | 1M input uses | Room left for the reply | Requests needed |
|---|---|---|---|---|
GPT-5.6 Sol / Terra / LunaMost room OpenAI · 1,050,000 | 1,050,000 | 95.2% | 50,000 | 1 |
Gemini 3.6 Flash / Flash-Lite Google · 1,048,576 | 1,048,576 | 95.4% | 48,576 | 1 |
Claude Opus 5 / Sonnet 5 Anthropic · 1,000,000 | 1,000,000 | 100.0% | 0 — no room for a reply | 1 (with zero output) |
Claude Haiku 4.5 Anthropic · 200,000 | 200,000 | n/a — 5x over | n/a | 5 |
Two models technically "fit" a bare 1M prompt and still cannot answer it: Claude Opus 5 and Sonnet 5 have a 1,000,000-token window shared with output, so a 1M-token prompt leaves zero tokens for the response. Each model also caps a single reply separately — 128,000 on GPT-5.6 and Claude Opus/Sonnet 5, 64,000 on Haiku 4.5, 65,536 on both Gemini models — which on GPT-5.6 means the 50,000 tokens of remaining window, not 128,000, is your real ceiling. Check your own document with the Context Window Checker.
1M in, 2K out — priced on every model
Fixed profile: 1,000,000 input tokens and a 2,000-token response. Because 1M is far above OpenAI's 272,000-token threshold, the three GPT-5.6 rows use long-context rates (input doubled, output 1.5x) while every other model uses standard rates.
| Model | Input rate applied | Input cost | Output cost | Per request | Per 1,000 requests |
|---|---|---|---|---|---|
GPT-5.6 Sol OpenAI · flagship (promo, long-context) | $8.00 / 1M | $8.0000 | $0.0600 | $8.0600 | $8,060 |
GPT-5.6 Terra OpenAI · mid (long-context) | $4.00 / 1M | $4.0000 | $0.0360 | $4.0360 | $4,036 |
Claude Opus 5 Anthropic · flagship | $5.00 / 1M | $5.0000 | $0.0500 | $5.0500 | $5,050 |
Claude Sonnet 5 Anthropic · mid | $2.00 / 1M | $2.0000 | $0.0200 | $2.0200 | $2,020 |
Claude Haiku 4.5 Anthropic · budget | $1.00 / 1M | $1.0000 | $0.0100 | $1.0100 | $1,010 |
Gemini 3.6 Flash Google · mid (promo rate) | $0.75 / 1M | $0.7500 | $0.0075 | $0.7575 | $757.50 |
Gemini 3.5 Flash-LiteLowest Google · budget | $0.30 / 1M | $0.3000 | $0.0050 | $0.3050 | $305 |
GPT-5.6 Luna OpenAI · budget (long-context) | $0.40 / 1M | $0.4000 | $0.0036 | $0.4036 | $404 |
Computed at 1,000,000 input and 2,000 output tokens 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 — both surcharges are proportional and would scale with the base rate. Cheapest-to-dearest spread is 26.4x, and output is under 1% of the bill on every model because the workload is almost entirely input. Caching, batch and taxes excluded.
At 1M, OpenAI's "cheaper" flagship costs more than Opus 5.
GPT-5.6 Sol lists at $4.00 per 1M input — a dollar under Claude Opus 5's $5.00. Every headline price table on this site shows Sol as the cheaper flagship, and for ordinary prompts that is correct.
A 1,000,000-token prompt is not an ordinary prompt. It is 3.7x over OpenAI's 272,000-token threshold, which doubles the input rate to $8.00 and lifts output by 1.5x to $30.00. Anthropic applies no equivalent surcharge, so Opus 5 stays at $5.00. The result is that a single 1M-token request costs $8.00 of input on Sol versus $5.00 on Opus 5 — the comparison reverses purely because of the size.
Chunking undoes it. Four requests of 250,000 input tokens each stay under the threshold and bill at the standard $4.00/1M, so the same million tokens cost $4.00 instead of $8.00 — exactly half. You give up cross-chunk attention and add three round trips, which is a real trade-off; but if the task is per-chunk extraction, summarisation or classification, the saving is free.
The general rule: 1M is a billing unit, not a request unit. Providers price per token, so a million tokens costs the same whether it arrives in one call or ten — except where a long-context surcharge makes the size itself expensive. Compare both shapes before assuming the biggest window is the best value.
How fast do you actually burn 1M input tokens?
Almost nobody sends a million tokens at once. This is what 1M input tokens means in the shapes real systems send — useful for sanity-checking a monthly line item.
| Workload | Input tokens per request | Requests in 1M input | What it looks like |
|---|---|---|---|
Support chatbotMost common Short history, short reply | 2,000 | 500 | Half a day of tickets for a small team. |
Document summarisation Retrieved chunks + instructions | 60,000 | 17 | A batch job, not a conversation. |
Code assistant Repo slice + file context | 30,000 | 34 | A few days of active development. |
Agent loop Tool results re-sent each step | 80,000 | 13 | One or two long autonomous runs. |
Long-context analysis Whole document in one prompt | 400,000 | 3 | Already inside the surcharge zone on OpenAI. |
Single full window Everything at once | 1,000,000 | 1 | Possible on only 5 of 8 tracked models. |
Request shapes are the presets used by the API Cost Calculator, rounded up to whole requests. Real traffic mixes shapes; use the calculator with your own input/output split and daily volume rather than extrapolating from this table.
Convert any token budget to words and pages
1M is the default; drop to 100K or 1K, or type your own figure. CJK capacity is shown separately because word counts do not apply.
1M-token questions
How many words is 1 million tokens?
About 750,000 English words, using the common 0.75 words-per-token average. The realistic range for plain prose is 714,000 to 800,000 words. Technical English gives less — roughly 625,000 to 714,000 words — and code, JSON or HTML far less again, because markup and punctuation consume tokens that carry no words.
How many pages is 1M tokens?
About 1,500 pages at 500 words per page for ordinary prose. Structured content collapses the page count: roughly 940 pages of source code, 800 pages of JSON and 730 pages of raw HTML.
Can you send 1 million tokens in a single request?
Only on 5 of the 8 models tracked here. GPT-5.6 Sol, Terra and Luna accept 1,050,000 tokens and leave 50,000 for the reply; Gemini 3.6 Flash and Flash-Lite accept 1,048,576 and leave 48,576. Claude Opus 5 and Sonnet 5 have a 1,000,000-token window, so a 1M-token prompt leaves zero room for a response, and Claude Haiku 4.5 needs five separate calls.
Does 1M tokens trigger OpenAI's long-context pricing?
Yes. OpenAI applies higher long-context rates above 272,000 input tokens on GPT-5.6 models, where input doubles and output rises 1.5x. GPT-5.6 Sol therefore bills $8.00 per million input tokens for a 1M-token prompt instead of the headline $4.00 — more than Claude Opus 5 at $5.00. Splitting the same million tokens into four 250,000-token chunks restores the standard $4.00 rate and halves the bill.
How much does 1 million input tokens cost?
For 1,000,000 input tokens plus a 2,000-token response, the cost runs from $0.3050 on Gemini 3.5 Flash-Lite to $8.0600 on GPT-5.6 Sol at its long-context rate — a 26.4x spread. Input dominates the bill at this shape: output is under 1% of the total on every model.
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, and the same text can meter several percent differently across OpenAI, Anthropic and Google. Cost figures multiply fixed token volumes by published per-1M rates taken from official provider pricing pages and re-verified weekly. The OpenAI long-context surcharge is applied because 1,000,000 input tokens exceeds the 272,000-token threshold on GPT-5.6 models; it is not applied to Anthropic or Google, which publish no equivalent tier. Caching, batch and tiered discounts are excluded throughout.
Official sources: OpenAI model docs, Anthropic pricing, Google Gemini API pricing. Always confirm the invoice before making purchasing decisions.