In AI, context is everything an AI chat such as ChatGPT, Claude or Gemini has in front of it at this moment: your question, the conversation so far, attached files and the provider's own instructions. The context window is the upper limit on how much of that fits in at once. When it is full, the AI has to summarise or drop older parts, and that is exactly when it seems to have forgotten something.
I work with Claude every day. My fourteen AI employees run on Claude Team, some also on open models or through the API. The figures in this article come from the help pages of Anthropic, OpenAI and Google, retrieved on October 26, 2026. I collect all the basics about AI at work on my AI Basics page.
What context means in AI
Behind every AI chat sits a language model, a Large Language Model or LLM for short. A language model does not remember anything between two answers on its own. For each answer it reads again whatever it is given. That is the context.
According to Anthropic and OpenAI, the context holds more than you see in the chat window:
- Your messages and the AI's answers, the whole conversation so far.
- Attached files and images, such as a PDF price list.
- The provider's instructions, known as the system prompt. These run invisibly before your first message.
- Tools and their results, for example from a web search or from a connected service such as Google Drive.
- Memories, if you have switched the memory feature on.
The difference from training knowledge matters. Before release, the model learned from a huge amount of text. That knowledge ends at a cut-off date and is the same for every user. The context, by contrast, is your desk for this one conversation. Anthropic calls the context window a "working memory" for that reason.
An example: you upload the price list of your painting business and ask for a quote for 80 square metres. Price list, question and answer are now in the context. If you open a new chat tomorrow and ask for the next quote, the AI no longer knows the price list. Unless, that is, you upload it again or have stored it permanently, for example in a project.
What a context window is
The context window is the maximum amount of context a model can process at once. It is measured in tokens. In AI, a token is a small piece of text, sometimes a whole word, sometimes just a syllable or a punctuation mark; it has nothing to do with crypto tokens. As a rough guide, Anthropic says 200,000 tokens are about 500 pages of text.
Two things surprise beginners: the AI's answer counts too, and the space for your own text is smaller than the whole window, because the system prompt, tools and memories take up room as well. OpenAI says so explicitly below its pricing table.
Reasoning models are models that think longer before they answer. This is how large the windows are according to the providers (as of October 26, 2026):
In Claude chat, the newest models Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 as well as Opus 5 and Sonnet 5 have 1 million tokens each, other models 500,000 or 200,000. The help page gives no figure for the free Claude plan. Which Claude models exist and what I use each one for is in Claude Models: Opus, Sonnet, Haiku, Fable.
A bigger window does not automatically mean better answers, though. Anthropic writes: "As token count grows, accuracy and recall degrade, a phenomenon known as context rot." The fuller the context, the more likely the model is to miss details. What is in the context matters just as much as the size of the window.
What happens when the context window is full
Anthropic describes this most precisely for Claude. On paid plans with code execution turned on, Claude automatically summarises older messages once a conversation approaches the limit, so it can carry on. Code execution means Claude may run small programs and create files for you. You find the switch in the settings under "Capabilities" as "Code execution and file creation"; on Free, Pro and Max it is on by default according to Anthropic. You may sometimes see Claude "organizing its thoughts" while this happens. According to Anthropic, the full history stays saved, but such a long conversation uses more of your usage allowance.
For ChatGPT and Gemini, the help pages I read give the size of the window but do not describe in detail what happens when it is full. Google writes that in an ongoing conversation Gemini has to re-read the conversation history each time, so long chats need more computing power.
So long chats get worse for two reasons: summaries lose details, and accuracy drops as the context fills up.
In Claude Code, Anthropic's coding tool, this summarising runs as a separate step that you can also trigger yourself.
Context, memory and projects: the difference
These three terms are often mixed up:
- Context is the working memory for a single conversation.
- Memory is a feature where the tool remembers things from earlier chats and brings them into new ones. These memories also end up in the context and take up space there. How this works in Claude is shown in Claude Memory: Check It and Use It.
- Projects are places for files and instructions that apply in every chat of the project. According to Anthropic, Claude uses a retrieval method to load only the parts of a project that match the question. How I set up projects is in How to Use Claude Projects Right.
Why I prefer to keep important knowledge in files rather than in the chat is explained in Why ChatGPT Forgets and AI Employees Don't. And if you run a model on your own computer, you set the context length yourself when you load it, as shown in Setting Up Local AI: Ollama to SGLang.
An example from everyday office work
A made-up but typical example: the owner of a small trade business has the AI write quotes in a single chat one morning. She uploads the price list at the start, then come customers A, B and C. On the fourth quote, the AI states the wrong hourly rate and carries over a travel charge from the first quote.
That is no accident. The price list sits at the very beginning of the conversation, three other customers lie in between, and the beginning may already have been summarised. A better setup: one chat per customer and the price list in a project, so it is freshly included in every chat.
What to watch out for
- One topic, one chat. When you switch tasks, start a new chat.
- Store important things permanently. Rules that should always apply belong in project instructions or files, not in message 37 of a long chat.
- Attach only what you need. A 200-page PDF fills the window even if you only need one page. How to still get long documents summarized well is covered in Summarize Texts and PDFs with AI.
- Switch off tools you don't need. Anthropic points out that connected services and web search cost a lot of tokens.
Frequently asked questions
What does context mean in AI?
Context is everything the AI has in front of it for its next answer: your messages, its previous answers, attached files, the provider's instructions and, if enabled, memories. Whatever is neither in the context nor part of its training knowledge, the AI does not know.
How large is ChatGPT's context window?
According to OpenAI, it depends on the plan and the model: 27,000 tokens on the free plan for the Instant models, 54,000 on Go and Plus, 128,000 on Pro. The Reasoning models have 256,000 tokens on Go and Plus and 400,000 on Pro (as of October 26, 2026).
How large is Claude's context window?
On paid plans, according to Anthropic, the newest models as well as Opus 5 and Sonnet 5 have 1 million tokens in chat, other models 500,000 or 200,000. Anthropic gives no figure for the free plan on that page.
What happens when the context window is full?
On paid plans with code execution turned on (settings, "Capabilities"), Claude automatically summarises older messages and carries on. If an answer suddenly no longer fits your early instructions, a new chat with a short summary is usually the best fix.
Is a bigger context window always better?
No. More space helps with long documents, but according to Anthropic accuracy drops as more tokens sit in the context. A tidy, short context often gives better answers than a full one.
Where to go from here
Open your longest chat and ask yourself whether the last task really still belongs there. If not, ask for a summary of where things stand and continue in a new chat. If you are just getting started with AI, read AI for Beginners: What to Know First first. Why an AI sometimes makes things up, especially in long chats, is explained in AI Hallucinations: Why AI Makes Things Up. You will find all the technical terms around AI briefly explained in AI Terminology: A Plain-English Glossary. How others organise their chats and AI employees is something we discuss in my German-speaking community Claude Practitioners.
