Claude Context Window Size: What Fits and What Carries Over
Claude’s context window is the information available for one response. Learn which limit applies to your session, what consumes that space, and what to save elsewhere.
Updated 5 min read
Claude’s context window size is the amount of information it can work with while producing a response, measured in tokens, the small pieces of text a model processes. The applicable capacity depends on the model and the product you use, so check the documentation for your actual session instead of relying on one universal Claude limit.
A large window does not tell you which project facts will reach the next conversation. Tools like fenbs keep current requirements and progress outside the chat, so you can bring back the information needed for the next task. This guide explains capacity; the context engineering collection covers other ways to prepare useful working material.
Which capacity applies to the Claude you are using?
First identify the product: Claude chat, Claude Code, or an application using the API, the direct interface for sending requests to a model. Then identify the selected model. A number advertised for an API model does not by itself establish what a particular app will let you attach or retain.
Use Anthropic’s model context documentation (opens in a new tab) for model capacities. For Claude chat, consult the usage and length limits explanation (opens in a new tab), which distinguishes conversation length from usage over time. Both can matter during the same afternoon, but they answer different questions.
Record the model and app when investigating a problem. “My research chat cannot accept this material” is a different report from “our application request exceeds the selected model’s window.” If you are comparing ordinary chat with a project’s knowledge collection, use the separate Projects context window guide.
What takes up the available space?
For API requests, Anthropic documents that instructions, messages, tool information, supplied documents, and the generated response all contribute to context. That makes the latest question only one part of the total. A short request can arrive after a long sequence of answers and tool results.
| Material | Example | Question to ask |
|---|---|---|
| Instructions | Project conventions and response requirements | Does this apply to the current task? |
| Conversation | Earlier drafts and corrections | Which version is approved? |
| Evidence | Documents, excerpts, and tool results | Which passages support this answer? |
| New response | A report or proposed implementation | How much output is actually needed? |
This is why pasting a large document and requesting another large document deserves planning. Leave room for the answer and the evidence it needs. Do not treat an input estimate as permission to fill every available token with background material.
Why is there no reliable page count?
A page is a visual layout; a token is a processing unit. A sparsely written slide, a dense contract page, and a screenshot of a table are different kinds of material. Their usefulness also differs: a readable heading may provide better guidance than an entire appendix unrelated to your question.
Avoid promising that a given window will hold a particular number of books or PDFs. Even a consistent page count would not establish that the relevant material was extracted or considered. Ask Claude to identify the source and section it used, then inspect that location yourself.
Say you are preparing a supplier comparison. You may need the warranty clauses and delivery conditions from several proposals, not every product photograph and installation appendix. Keep the originals available, but give the immediate question a clear evidence boundary. Our AI context window explanation describes the general concept beyond Claude.
How can you check consumption without guessing?
Use a measurement appropriate to the interface. For an API integration, Anthropic provides a token counting endpoint (opens in a new tab) for supported inputs before a request. The response’s usage information reports consumption afterward. Check the endpoint’s supported input types before assuming it measures your entire workflow.
For Claude Code, use the context information described in its context management guidance (opens in a new tab). Inspect what occupies the session before removing useful instructions. An API estimate for a separate prompt does not reveal a consumer chat’s hidden instructions or exact remaining space.
- Note the app, selected model, task, and any exact warning shown.
- Inventory the sources, old drafts, instructions, and tool output involved.
- Use the app’s available indicators or the API’s documented measurement tools.
- Reduce one unnecessary category and check whether the same task can proceed.
Keep your investigation reproducible. If a smaller source set works, write down which material you removed. That gives you a practical operating habit without pretending you measured something the interface does not expose.
Does fitting everything guarantee a better answer?
No. Available capacity and reliable use are different tests. Anthropic notes that more context is not automatically better. For your own work, check whether the answer follows the relevant rule and cites the right evidence rather than treating successful submission as proof of understanding.
An old draft can disagree with an approved revision even when both fit. Label the current version and describe what replaced the earlier one. If two sources still conflict, ask for the conflict to be reported before Claude produces a recommendation based on either.
A useful review question is: “Which requirement most affects your answer, and where did it come from?” Compare that answer with the source. For readers choosing between Claude products, what Claude is provides the broader introduction without making capacity the only selection criterion.
What should you preserve beyond the active window?
Save the goal, the approved decisions, the current result, and the next unresolved action. Include source locations for details that should be checked again. You want enough information to resume responsibly, rather than a transcript that makes the next session repeat every rejected idea.
For the supplier example, save that the delivery date is confirmed, the warranty exclusion needs review, and the recommendation has not been approved. Keep the proposal version beside each conclusion. That makes a fresh session useful even if the earlier conversation is not part of its working material.
This is the practical role of fenbs project context: Rules and Where we left off give a connected assistant a current starting point. For repository work, the guide to context engineering for Claude Code applies the same selection principle to coding tasks.
Bring back the project facts that matter now
Keep approved decisions, current tasks, and source references in fenbs. Ask your connected assistant to read the project context, then open only the evidence needed for the next task. You can preserve the work without carrying every earlier draft into the conversation.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account, free for one person with no credit card, and add the project you need to continue.
- Record the current goal in the project, agreed choices in Decisions and rules, and the next action on its task.
- Connect Claude using the appropriate integration guide over MCP, a standard way for an AI app to use an outside tool.
- Ask the assistant to read Rules and Where we left off before working, and save checked results and remaining questions before stopping.
fenbs does not enlarge Claude’s context window or measure hidden chat consumption. If your immediate need is only to count a single API request, use Anthropic’s measurement tools. Use fenbs when the harder problem is keeping the right work available for the next session.