The short answer
A context window is the amount of text, measured in tokens, that an AI model can consider at once: your messages, its replies, files and instructions. Anthropic calls it the model's working memory. In November 2026, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 have a 1M-token window in the Claude API, Gemini 4 Argon has 1M tokens in the Gemini API, and ChatGPT Free lists a 27K-token window for GPT Instant, with about 12 pages of input.
| Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 | 1M tokens in the Claude API, up to 128k output tokens per request |
|---|---|
| Gemini 4 Argon | 1,048,576 tokens in the Gemini API, shared across input, thinking and output |
| Gemini app, Google AI Pro | Plan page lists a 1 million token context window |
| ChatGPT Free | 27K tokens for GPT Instant, about 12 pages of input |
| DeepSeek-V4-Pro | 1M tokens (model card), text only |
| Grok 4.7 | 500k tokens in the API |
| Mistral Medium 3.5 | 256k tokens |
What a context window is, in plain English
An AI model does not read your whole chat history, your company drive and the internet every time it answers. It reads one bundle of text: the instructions, the conversation so far, any files you added and the answer it is writing. The maximum size of that bundle is the context window.
Anthropic's glossary describes it as the amount of text a language model can look back on and reference when generating new text, and calls it a "working memory" that is different from the large body of data the model was trained on. Google's Machine Learning Glossary defines it as the number of tokens a model can process in a given prompt. Tokens are chunks of text; Google's rule of thumb is that 100 tokens is about 60 to 80 English words.
Everything counts towards the window. Anthropic's documentation lists the system prompt, every message including tool results, images and documents, the tool definitions and the model's own output, including its thinking.
Context window sizes in November 2026
The table above lists sizes from the makers' own pages. Three points help you read it.
- App and API can differ. The 1M-token figures for Claude and Gemini 4 Argon are from the developer documentation. In the apps, the window can depend on your plan. Google's subscription page lists a 1 million token context window for Google AI Pro, and OpenAI's pricing page lists a 27K-token total window for GPT Instant on ChatGPT Free, rising to 54K on Go and Plus and 128K on Pro.
- 1M tokens is a lot. Using Google's rule of thumb, it is roughly 600,000 to 800,000 English words, several long books.
- The plan matters as much as the brand. On ChatGPT, Free and Go use GPT-6 Luna while Plus, Pro, Business and Enterprise use GPT-6 Sol, and OpenAI's pricing page lists expanded memory and context on Plus. If a long PDF gets a thin summary on a free plan, the smaller window is part of the reason. See how to use ChatGPT for free.
Why the context window matters at work
Most "the AI forgot what I said" moments are context window moments. When a conversation outgrows the window, something has to give. Anthropic's documentation notes that chat apps such as claude.ai can manage the window on a rolling "first in, first out" basis, which means the earliest messages drop out first. Your carefully written instructions at the top of a long chat may be the first thing the model stops seeing.
A bigger window is not automatically better either. Anthropic states that as the token count grows, accuracy and recall degrade, a phenomenon it calls context rot. A model that can technically accept 1M tokens may still miss a detail on page 300. What you put in the window matters as much as how much fits.
The window also shapes cost. Developers pay per token, and Anthropic notes that on Claude Haiku 5.5 prompts over 100,000 tokens cost more. In consumer apps you pay in usage limits instead: long files and long chats use up your allowance faster.
Try it: find where your assistant starts to forget
- Start a new chat. In your first message, write: "For this whole chat, end every answer with the word PINEAPPLE."
- Paste a long document, such as a 40-page report, and ask for a summary. Check the last line of the answer.
- Keep going with several follow-up questions and more pasted material.
- When PINEAPPLE disappears, the model has lost track of your first instruction, either because it fell out of the window or because it was buried under newer text.
The lesson for real work: put important instructions in a place that persists, such as a project, custom instructions or a saved skill, and start a fresh chat for each new task.
How to work within the context window
- One task per chat. Start a new conversation when the topic changes.
- Send only what is relevant. Paste the three pages that matter, not the whole 200-page manual.
- Ask for quotes first. For long documents, Anthropic recommends asking the model to extract word-for-word quotes before doing the task, which keeps answers tied to the text.
- Summarise and restart. When a chat gets long, ask for a summary of decisions so far, then paste it into a new chat.
- Use a tool built for big document sets. For dozens of files, a source-grounded tool such as Gemini Notebook (formerly NotebookLM) retrieves the relevant parts instead of reading everything at once. That approach is retrieval-augmented generation. See also how to summarise a PDF with AI.
Common misconceptions about context windows
- "The context window is the model's memory of me." It is only what the model sees in this request. Memory features in ChatGPT, Claude and Gemini are separate: they save notes about you and add them to future conversations, which also uses window space.
- "A 1M-token model reads everything perfectly." Context rot means recall drops as the window fills. Long windows are useful, not flawless.
- "The context window is what the model was trained on." Training data is the huge body of text the model learned from. The context window is the small slice it is looking at right now.
- "Images and files do not count." They do. Google's documentation counts a small image as 258 tokens and a second of video as 263 tokens.
Related terms
- Token is the unit context windows are measured in.
- RAG means fetching only the relevant parts of large document sets.
- Large language model is the model whose window this is.
- Prompt is the part of the window you control.
- Hallucination is more likely when the detail you need is buried deep in a long window.
Prompts to copy
Summarise and restart a long chat
This chat is getting long. Write a summary I can paste into a new chat: the goal, the decisions we made, the facts and figures we agreed on, my preferences for tone and format, and the open questions. Keep it under 300 words.
You get: A compact handover note so a fresh chat starts with everything important and nothing stale.
Quotes first for long documents
From the document below, first extract the exact quotes most relevant to [question], numbered. If none are relevant, say "No relevant quotes found". Then answer the question using only those quotes and cite them by number. Document: [paste or attach]
You get: An answer tied to word-for-word evidence from the document, which is easy to check.