Context windows explained: how much fits in 128K, 256K and 1M tokens

A context window is the number of tokens a model can work with at once, counting everything: instructions, documents, the conversation so far and the answer it writes. In English, 128K tokens hold roughly 77,000 to 102,000 words, 256K about twice that, and 1M tokens roughly 600,000 to 800,000 words. Most current flagship API models have windows of 1M tokens; consumer chat plans often give you less.

What fits in a window

Using Google’s rule of thumb that 100 tokens are about 60 to 80 English words, and 500 words to a page:

Window	English words	Pages (500 words)	Example
27K	16,000 to 22,000	32 to 43	A long report
54K	32,000 to 43,000	65 to 86	A short book or a large contract
128K	77,000 to 102,000	154 to 205	A full-length novel
256K	154,000 to 205,000	307 to 410	A long novel, or several contracts
1M	600,000 to 800,000	1,200 to 1,600	A shelf of books, or a mid-sized codebase

Code, tables and non-English text use more tokens per word, so the same window holds less of them. For your own document, the context window checker counts it and compares it with every model.

Windows of current API models

Model	Context window	Maximum output
GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna	1.05M	128K
Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5	1M	128K
Gemini 4 Argon, 3.8 Flash, 3.5 Flash-Lite	1,048,576 (input and output together)	65,536 by default
DeepSeek-V4-Pro, V4.1-Flash	1M	up to 384K
Grok 4.3	1M	not stated
Grok 4.7	500K	not stated
Mistral Large 4	1M	not stated
Mistral Medium 3.5, Small 4	256K	not stated

Google notes that Gemini 4 Argon’s 1,048,576 tokens are a single limit for input and output together, and that its output can go beyond the default 65,536 tokens with long decoding.

Chat apps get smaller windows

The window in a chat app is set by the plan, not by the model’s API limit. OpenAI publishes its figures: ChatGPT Free has a 27K window with GPT Instant, Go and Plus 54K, and Pro 128K; with GPT Reasoning, Go and Plus get 256K and Pro 400K. OpenAI adds that system instructions, memory and the model’s reasoning share that window, so the space for your own text is smaller. Anthropic and Google do not publish per-plan window sizes on their pricing pages.

Why the answer needs room

The model writes its answer into the same window that holds your input. If a document fills 990,000 tokens of a 1M window, only 10,000 tokens are left for the reply. When you plan, add the answer you expect to the document size. The context window checker does this with a reserve you can set.

A bigger window is not always better

• Cost: every token you send is billed, every time. Sending a 500,000-token document to Claude Sonnet 5.5 costs $1 in input per request; sending it 100 times costs $100, unless prompt caching applies.
• Price tiers: some models charge more above a threshold. OpenAI’s GPT-6 prices double on input above 272,000 tokens, xAI bills the whole request at a higher rate from 200,000 prompt tokens, and Claude Haiku 5.5 costs more above 100,000.
• Attention: a model given a huge context has to find the relevant part within it. For questions about one section of a long document, sending only that section is often cheaper and gives a more focused answer.

When a document is too long

Split it into parts that each fit with room to spare, summarize or question each part, then combine the results. Or index the document and send only the passages relevant to each question, the approach known as retrieval. The checker tells you how many parts a document needs for each model.

Sources

• Google, Understand and count tokens and Gemini 4 Argon.
• OpenAI, models, pricing and ChatGPT plans.
• Anthropic, models overview and pricing.
• DeepSeek, models and pricing; xAI, models; Mistral, models overview.

All pages read on October 26, 2026.

Free toolWill my whole document fit?Check a document