Context window explained simply: how much AI remembers in a conversation

What the context window is

For each answer, an AI chatbot reads not only your last message but as much of the conversation so far as fits into its context window. Anthropic compares the context window to working memory: it determines how much content the AI can process and keep track of at once in a single chat. Google calls it reading capacity.

Its size is measured in tokens, the small pieces of text an AI splits every text into. We have a separate explainer on that.

It is not only your text that goes into the window. OpenAI writes that ChatGPT also uses it for system instructions, tools, saved memories and internal processing, such as thinking and composing the answer. The space for your input is therefore smaller than the whole window. Anthropic also always reserves part of it for the answer.

How large context windows are today

ChatGPT: according to OpenAI's German pricing page, the quick Instant level has a total window of 27,000 tokens on Free, 54,000 on Go and Plus, and 128,000 on Pro. For thinking answers it is 256,000 on Go and Plus and 400,000 tokens on Pro; on Free the value varies. OpenAI translates this into pages: with Instant, about 12 pages of text fit as input on Free, about 40 on Plus and about 250 on Pro.

Gemini: according to Google's help pages, it is 32,000 tokens without a subscription, 128,000 with Google AI Plus and one million tokens with AI Pro or AI Ultra. According to Google, one million tokens equals up to 1,500 pages of text.

Claude: on paid plans, Anthropic says the newest models such as Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 have a window of one million tokens when chatting. Other models have 500,000 or 200,000 tokens; 200,000 tokens are about 500 pages, according to Anthropic. The help page reviewed gives no figure for the free plan.

Important: the large windows are mostly only available on paid plans. If you chat for free, ChatGPT and Gemini have a much smaller memory per conversation.

What happens when the window is full

The vendors handle this differently. Google warns that when the context window is exceeded, Gemini may not fully take content into account, may miss connections or leave out details, especially when a question refers to many details spread across large files.

On paid plans, Claude automatically summarises earlier messages when a conversation approaches the limit, provided the code execution feature is turned on. Anthropic writes that this lets conversations continue in most cases, but that such long chats use more of your usage limit. Otherwise, the only option is a new chat.

Even a large window does not mean the AI pays equal attention to everything in it. Researchers at the company Chroma tested 18 language models in July 2025 and found that performance drops as input length grows, often unevenly and in surprising ways. Our assessment: enter important things briefly and specifically rather than relying on a huge window.

Context window and memory are two different things

Many chatbots also offer a memory feature. In ChatGPT it is called Memory. According to OpenAI, ChatGPT can use saved memories and, if turned on, information from past chats to tailor future answers. These memories are stored separately from the chat history and, according to OpenAI, do not retain every detail of every conversation.

The difference: the context window is what the AI has in front of it in this one conversation. Memory is a kind of note across several conversations, for instance that you are vegetarian. If an AI has forgotten something, it may have dropped out of the context window or never been saved as a memory. But the AI can also overlook information that is still in the context window: according to a study by Chroma, language models become less reliable the longer the input gets.

How to use the context window wisely

A few simple rules can be drawn from the vendors' advice. Start a new chat for a new topic. Paste only the pages of a document you actually need. For inputs that are too large, OpenAI advises shortening the text, removing repetition, splitting it into smaller parts or summarising it first.

If a conversation has become very long, you can ask the AI to summarise the key points and use that summary in a new chat. For Claude, Anthropic also recommends projects, where only the relevant parts of your files are loaded into the window, and turning off tools you do not need, because they take up a lot of space.