It is the total amount of text Claude can hold in view at one time while it works on your request.

A context window is Claude’s working memory for a conversation. Everything that Claude can read and consider when it produces its next reply has to fit inside it. That includes the messages you have typed, the replies Claude has already written, any files or documents you have added, instructions you have saved, and the reply Claude is in the middle of writing right now.

The size of the window is measured in tokens, which are small chunks of text, roughly a few characters each. A bigger window means Claude can keep more material in view at once, such as a long report, several documents, or a very long conversation. A smaller window means it has less room, so older material must eventually be set aside.

flowchart LR
    A["Your messages"] --> E["Context window"]
    B["Claude's earlier replies"] --> E
    C["Files and documents you add"] --> E
    D["Saved instructions and tools"] --> E
    E --> F["Claude writes the next reply"]
        

A few simple words appear again and again in this topic. Here they are in plain English.

Context windows are counted in tokens, not in words or pages.

Before a language model reads text, it breaks the text into tokens. A short common word is often one token. A longer or rarer word may be split into several. Punctuation, spaces, and numbers also count. As a rough guide for ordinary English, one token is about four characters, so a typical page of text uses a few hundred tokens. Other languages, code, and unusual formats can use more tokens for the same amount of meaning.

Anthropic’s Help Center gives a helpful sense of scale: a window of two hundred thousand tokens is about five hundred pages of text or more. A window five times larger would hold about five times as much material. You do not need to count tokens yourself, but knowing the rough scale helps you judge whether a task fits.

The size depends on the model, the product you use, and your plan.

Claude’s context window is not a single fixed number. Anthropic has increased it over time, and different models have different sizes. According to Anthropic’s Help Center article on the context window for paid plans, the size when chatting with Claude ranges between about two hundred thousand and one million tokens, depending on the model. Among the newest models in the lineup, Claude Fable 5.1, Claude Opus 5.5, and Claude Sonnet 5.5 are listed with a one million token window in chat. Some earlier models are listed with five hundred thousand tokens, and others with two hundred thousand.

The same article notes that sizes can differ in other products, such as Claude Code, which is built for software work. Some models have larger windows there, and certain options may need extra usage enabled. Because these figures are updated whenever new models arrive, treat the numbers in this guide as a snapshot and confirm the current table before depending on a particular size.

More things take up room than most people expect.

When you send a message, it is natural to think only of the words you just typed. But Claude reads far more than that each time. To reply well in a conversation, it takes into account the whole conversation so far, plus everything else that has been added to it. Every item takes up some of the window.

flowchart TD
    A["Context window as a container"] --> B["Conversation history"]
    A --> C["Files and documents"]
    A --> D["Instructions and memory"]
    A --> E["Tools and connectors"]
    A --> F["Thinking and the reply being written"]
    B --> G["Together they must fit within the limit"]
    C --> G
    D --> G
    E --> G
    F --> G
        

Each new message makes Claude re-read the whole conversation, so the load grows over time.

Claude does not keep a private diary between your messages. Each time you send something, the earlier conversation is provided again as part of the input, so that Claude can answer in context. This means a conversation of fifty messages is much heavier to process than a conversation of five. The window fills up gradually, and the work for each reply grows along with it.

This is also why long conversations can use more of your usage allowance. More text to read means more work per reply. If you notice your allowance draining faster in a very long chat, this is a common reason.

On paid plans Claude can manage this for you, in most cases.

According to Anthropic’s Help Center, for users on paid plans with code execution enabled, Claude automatically manages the conversation context. When a conversation approaches the window limit, Claude summarizes earlier messages to make room for new content. This allows conversations to continue indefinitely in most cases. Your full chat history is preserved, so Claude can still refer to it, even after earlier parts have been summarized. You may occasionally see Claude “organizing its thoughts” in a long conversation, and that is this management at work.

The same article adds some honest limits. Code execution needs to be enabled for the automatic management to work. In rare edge cases, such as a very large first message or a system error, you can still run into the window limit. And longer conversations that trigger automatic management use more of your usage limit.

flowchart TD
    A["Conversation grows"] --> B{"Is the window nearly full?"}
    B -- "No" --> C["Carry on as normal"]
    B -- "Yes, on a paid plan with code execution" --> D["Claude summarizes earlier messages"]
    D --> E["Space is freed for new content"]
    E --> C
    B -- "Rare edge case or feature off" --> F["You may hit the limit and need a new chat"]
        

A large window is a big help, but filling it to the brim has a cost.

It is tempting to think that if the window can hold a million tokens, you should paste in everything you have. Anthropic’s developer documentation gives a useful warning. A larger context window lets the model handle more complex and lengthy prompts, but more context is not automatically better. As the number of tokens grows, accuracy and recall can degrade, a phenomenon the documentation calls context rot. In other words, what is in the window matters as much as how much fits.

A helpful way to think about it is quality over quantity. A focused window containing exactly the material needed for the task usually serves you better than a window stuffed with loosely related content. This is why organizing your material and removing clutter is a smart habit, even when the window is huge.

The context window and memory are two different things.

People often mix these up. The context window is what Claude can see right now, within the current conversation. Memory, a separate feature you can turn on, lets Claude save useful details from your chats and bring them back in later conversations. One is short-term working space. The other is a notebook that carries across chats.

So when Claude “forgets” something from earlier in a very long chat, the reason is usually the window: older material was summarized or crowded out. When Claude “remembers” something in a brand new chat, that comes from memory or saved instructions, which are loaded into the new window at the start.

Projects use a clever trick to work with more material than fits in the window at once.

A project lets you keep many documents together as project knowledge. Anthropic explains that projects use retrieval-augmented generation, often shortened to RAG. In simple terms, instead of loading every document into the window each time, Claude looks up and loads only the relevant pieces for your question. This lets it work with larger amounts of information than would otherwise fit.

When Claude thinks before answering, that thinking needs room too.

Some models can reason through a problem before replying. According to Anthropic’s developer documentation, when thinking is used, the tokens involved count toward the context window while Claude is working. The documentation also explains that in later turns of a conversation, earlier thinking is generally not carried forward in the same way, which helps save room for the actual conversation. The details differ between the apps and the developer tools, so treat this as the general idea.

The practical lesson for everyday users is simple. Higher effort and deeper thinking give better results on hard problems, but they use more of your resources. If you are working with a very large document and also need very deep reasoning, you are asking for a lot from the window at once, so plan accordingly.

People who build with the Claude API meet the window more directly.

Developers who use the Claude API can see and manage the window more precisely. Anthropic’s documentation describes a few ideas that are useful to know even if you never write code.

Certain symptoms suggest it is time to tidy up or start fresh.

A few easy habits keep Claude sharp and your allowance healthy.

Let us follow an imagined office manager named Rohan as he works with a very large document.

Rohan used the window well. He picked enough room, shared only relevant material, worked in stages, and verified the outcome. That is the whole spirit of managing context.

Most window problems come from a few common habits.

A big window opens new possibilities and brings a few new things to watch.

Context windows have grown over time, and the numbers you read today may be out of date soon.

Early AI chat tools could hold only a few pages of text. Over the years, windows grew to hundreds of thousands of tokens and then to a million. Each new Claude model may bring a different size, and the same model can offer different sizes in different products. Some larger sizes may also need particular settings or extra usage to be enabled.

Because of this, the safest habit is to treat any specific number, including those in this guide, as a snapshot. When size matters to your work, check the Help Center article on the context window for paid plans, and the developer documentation if you build with the API, where the model comparison table lists sizes and other details.

flowchart LR
    A["Early tools: a few pages"] --> B["Hundreds of thousands of tokens"]
    B --> C["Up to a million tokens on newer models"]
    C --> D["Better management: summarizing, retrieval, compaction"]
        

Quick answers to the questions beginners ask most often.

It is the amount of text Claude can keep in view at one time while working on your request, including your messages, its replies, files, and saved instructions.

In tokens, which are small chunks of text. In everyday English a token is about four characters, and two hundred thousand tokens is about five hundred pages or more.

It depends on the model, the product, and your plan. On paid plans it ranges from about two hundred thousand up to one million tokens, with the newest models at the larger end. Check the Help Center article for the current table.

On paid plans with code execution enabled, Claude summarizes earlier messages to make room, so conversations can usually continue. In rare cases, such as a very large first message, you may still hit the limit.

Earlier material may have been summarized, or crowded out by a lot of other content. Restate important details in a recent message.

No. The window is the working space for the current conversation. Memory is a feature that saves details across chats, and what it brings back is loaded into the window.

Yes. Claude reads the conversation again for each reply, so longer chats take more work, and conversations that trigger automatic context management use more of your usage limit.

Not always. Larger windows are very useful, but accuracy and recall can weaken as they fill up, so focused, relevant material often works better.

Projects use retrieval, which loads only the relevant parts of your stored documents into the window, so they can work with more material than would otherwise fit.

Yes. Anthropic notes they are token-intensive, so keeping unused ones switched off leaves more room for your work.

Ask for a summary of key decisions, then start a new chat with that summary, or restate the key facts. Use one chat per topic where possible.

In the Claude Help Center article on the context window for paid plans, and in the developer documentation page on context windows for API details.

The context window is the amount of text Claude can keep in view at once, measured in tokens. It holds your messages, Claude’s replies, files, saved instructions, tools, and any thinking Claude does. On paid plans its size ranges from about two hundred thousand to one million tokens depending on the model, and the figures change as new models arrive. As a chat grows, each reply takes more work, and when the window nears its limit Claude can summarize earlier parts so the conversation can continue, though some detail may be lost. A bigger window is helpful, but a full one can weaken accuracy, so focused, relevant material is best. Projects use retrieval to load only what is needed, memory carries details across chats, and good habits such as one topic per chat keep Claude sharp and your usage healthy.