Level : Intermediate
Context window: how much text can an AI read at once?
The context window is the amount of text an AI model can take into account at one time. The tool above converts it into pages, novels and hours of meetings so you can compare models and check which ones accept your full document.
12 of 12 models accept this document in full (≈ 100,000 tokens).
Context window size (in A4 pages)
Vertical line: size of the chosen document. Orange bar: the document does not fit in full.
| Model | Windowtokens | A4 pages | Novelsof 300 pages | Meetingtranscribed hours | Chosen document | Remove |
|---|---|---|---|---|---|---|
| GPT-6.1 SolOpenAI | 1,050,000 | 1,575 | 5.3 | 88 | ✓ fits9.5% of the window | |
| DeepSeek V4.1 Flash 1017DeepSeek | 1,048,576 | 1,573 | 5.2 | 87 | ✓ fits9.5% of the window | |
| Gemini 4 ArgonGoogle DeepMind | 1,048,576 | 1,573 | 5.2 | 87 | ✓ fits9.5% of the window | |
| Claude Sonnet 5.5Anthropic | 1,000,000 | 1,500 | 5 | 83 | ✓ fits10% of the window | |
| GLM-5.3Zhipu AI (Z.ai) | 1,000,000 | 1,500 | 5 | 83 | ✓ fits10% of the window | |
| Qwen3.8 Max 0902Alibaba Cloud (Qwen) | 1,000,000 | 1,500 | 5 | 83 | ✓ fits10% of the window | |
| Grok 4.6xAI | 500,000 | 750 | 2.5 | 42 | ✓ fits20% of the window | |
| GPT-5OpenAI | 400,000 | 600 | 2 | 33 | ✓ fits25% of the window | |
| Kimi K2.6Moonshot AI | 262,144 | 393 | 1.3 | 22 | ✓ fits38% of the window | |
| Mistral Medium 3.5Mistral AI | 262,144 | 393 | 1.3 | 22 | ✓ fits38% of the window | |
| DeepSeek V3.2DeepSeek | 128,000 | 192 | 0.6 | 11 | ✓ fits78% of the window | |
| Llama-3.3-70B-InstructMeta | 128,000 | 192 | 0.6 | 11 | ✓ fits78% of the window |
Assumptions: 1 token ≈ 0.75 words; 1 A4 page ≈ 500 words; 1 novel = 300 pages; 1 hour of transcribed meeting ≈ 9,000 words (about 150 words per minute); code ≈ 10 tokens per line. These are orders of magnitude to compare models: text in other languages than English, tables or code fill the window faster.
The model's reply also takes up the window, so keep a margin if you expect a long answer. Context windows collected on October 22, 2026.
The working memory of a conversation
When you chat with an AI assistant, the model does not remember things the way a person does. With every message, it rereads everything it is given: its instructions, the conversation so far, attached documents and your new question. The context window is the maximum amount of text it can take into account at once. It is measured in tokens, chunks of words. Anything beyond it is rejected, cut off or summarized, depending on the app.
Think of it as a desk: the bigger it is, the more documents you can spread out at the same time. But when the conversation ends, the desk is cleared.
Not to be confused with long-term memory
Some assistants offer a “memory” feature that keeps information from one conversation to the next: your name, your job, your preferences. That is a different mechanism. Those details are stored separately and slipped into the context window at the start of a new chat. Long-term memory decides what gets recalled; the context window limits how much text can be read at the same time.
How to read the conversions
A token count on its own means little. The tool above translates it into concrete equivalents: A4 pages, 300-page novels and hours of transcribed meetings. These conversions rely on averages (words per page, speaking pace, tokens per word) that vary with language, layout and density. A spreadsheet, code or a text in French or Spanish will not fill the window at the same rate as an English novel.
So treat these equivalents as ballpark figures for comparing models, not exact measurements. The bars show the gaps at a glance; then pick a sample document to see which models can accept it in full. For a specific text, the token estimator gives a closer estimate.
A large window does not guarantee careful reading
Fitting a document into the window does not mean the model gives every line equal attention. Researchers have described an effect nicknamed “lost in the middle”: models often retrieve information placed at the beginning or end of a long text better than information buried in the middle. Recent models are improving on this, but caution is still wise.
The longer the context, the higher the risk of something being missed or mixed up, and the more each request costs, since all those tokens are billed on every call. The AI agent cost calculator lets you measure that effect.
Tips for working with long documents
- Split it up: process a long document chapter by chapter or section by section, then ask for a summary of the partial answers.
- Summarize as you go: in a long conversation, ask for a summary now and then and start a fresh chat from it.
- Put the key parts in the right place: place your question and main instructions after the document, or repeat them at the end.
- Use RAG: retrieval-augmented generation means storing your documents in a searchable database and giving the model only the passages relevant to each question. Think of a librarian who brings you the right pages instead of the whole library. It is a common approach for agents that work on large document collections.
To compare other model specs, use the model comparison tool. The Go further learning path and the glossary cover these ideas in more depth.
Frequently asked questions
What happens when the context window is full?
It depends on the app. Some reject the document, others cut the text or drop the earliest messages in the conversation, and others summarize the history automatically. Either way, part of the information is no longer read as it was.
Is the context window the same as ChatGPT's memory feature?
No. The context window is the working memory of a single conversation and is cleared when it ends. The long-term memory offered by some assistants stores a few facts about you separately and feeds them into new conversations.
Are the page conversions exact?
No, they are ballpark figures based on averages of words per page and tokens per word. Dense or technical text, or text in languages other than English, fills the window faster than an airy English novel. They are mainly useful for comparing models with each other.
Does the model's answer count toward the context window?
Usually, yes. The window has to hold both the text you send and the reply, and some models also set a separate cap on output length. Leave some room if you expect a long answer.