A context window is the amount of information a language model can work with while producing a response. It includes the material supplied to the model and space for the answer it generates. Think of it as the papers currently spread across a desk: useful working material for the present task, rather than everything stored in the building. A larger desk can hold more documents, but it does not guarantee a correct reading of every detail.
| Term | What it describes |
|---|---|
| Context window | Information available for the current response, including generated output |
| Output limit | The maximum size of a generated answer |
| Training data | Material used to train a model, separate from the current conversation |
| Compaction | Replacing older conversation material with a summary |
Anthropic's official context window documentation, read on 4 November 2026 at 22:03 IST, describes this as a model's working memory. The explanation is useful beyond a single app. It helps answer why a long conversation, several attachments and a detailed instruction can compete for space, even when each individual message looks manageable on screen.
What counts towards an AI context window?
In Claude's API, the request can include system instructions, conversation messages, tool definitions and tool results. Images and documents also contribute to the context budget. The generated response counts too. This means the visible question you just typed is only one part of what the model may need to process. The application decides what other material accompanies it.
Consider an original example: a student supplies a course outline, several chapters and a request for a revision plan. The outline and chapters take up input space. Earlier discussion may take up more. The revision plan requires output space. Counting only the final request, such as “make a timetable”, would miss most of the information involved in that interaction.
Context is measured in tokens, the units the model processes. A token count is not a reliable substitute for a word count. For a document in Hindi, English or a mixture of languages, avoid assuming that a fixed number of pages will always consume the same space. Formatting, content and the model's processing rules make a document's actual token count more useful than a page based guess.
Context window versus output limit: what is the difference?
The context window describes the overall working space. The output limit describes how much the model can generate in one response. A system can therefore accept a substantial document collection without being able to reproduce an equally substantial answer in one turn. Asking for an extremely long report is a different demand from asking a short question about a long report.
Here is a hypothetical illustration, not a specification for a particular service. Suppose a system has a total budget of 100 units and your input uses 90. The remaining capacity is constrained even if the separate output setting permits 20 units. Real products use tokens and their own rules, but the arithmetic explains why input size and answer length must be considered together.
Anthropic provides a token counting endpoint for developers to estimate input before sending a message. Its supported inputs include system prompts, tools, images and PDFs. For ordinary app users, the practical lesson is to consult that product's own limits rather than treating an API specification as a promise about every subscription or chat interface.
Is context the same as memory or training knowledge?
No. Training is separate from the information supplied in the current request. Likewise, an application keeping a conversation in your history does not establish that the entire conversation will be passed to the model on every later turn. Anthropic notes that chat interfaces can manage context on a rolling basis. The contents you can scroll back to and the contents currently being processed are different questions.
Long conversations can also be condensed. Anthropic's compaction overview describes replacing older turns with a summary. Summaries preserve a shorter account of earlier work, rather than retaining every sentence verbatim. If a later decision depends on exact wording, supply the relevant passage again and ask for a response grounded in that passage.
How can you make better use of limited context?
Use clear instructions, identify the documents that matter and remove irrelevant material. Anthropic's context engineering guidance warns that more context does not automatically improve recall or accuracy. That is a provider's stated limitation, not a benchmark comparison between services. A large advertised capacity should never be read as a guarantee that every attached detail will be interpreted correctly.
An original prompt for comparing two notices is: “Use only Notice A and Notice B below. List changed eligibility conditions, quote the relevant sentence from each, and say when either notice is silent.” Its purpose is to define the evidence and comparison task. It cannot guarantee correctness, so check the quotations against both originals before acting on the result.
For related concepts, read our RAG explainer, what an AI agent does and how Claude Projects organises instructions and files. They help distinguish finding information, carrying out tasks and organising a workspace from the limited context used for an individual response.
