Claude Opus Context Window: Versions, Capacity, and Restarts
Check the context window of the exact Opus version you use, then separate request capacity from app usage limits. A reviewed work record makes a restart easier when either limit interrupts a session.
Updated 5 min read
To check the context window of Opus 5, use Anthropic's specifications for the exact model and the application through which you access it. The context window describes how much material a request can work with; it is different from the usage allowance that controls how much you can use an app over time.
These limits call for different responses. Shortening a document can help a request fit, but it does not reset an account allowance. fenbs keeps a reviewed record of the work outside the chat, giving the next session a clear starting point when you need to restart or switch tools.
Which Opus specification should you check?
Anthropic provides separate model pages for Claude Opus 5 (opens in a new tab) and Claude Opus 4.8 (opens in a new tab). Match the selected version to its exact identifier, such as claude-opus-5 or claude-opus-4-8. Read the current context and output specifications there instead of assuming that all Opus models have identical limits.
If you only see Opus in an application, check which version that label represents. Record the model and route with a failed request so you can reproduce it. Avoid using a screenshot of an old model picker as evidence for what your current session supports.
A useful comparison has a specific question: can this setup analyze the selected sources and produce the answer I need? The general Sonnet versus Opus guide addresses choosing between families. This page focuses on identifying capacity and preserving work when the current session cannot continue.
What changes when you use Opus through a different app?
A model's API specification and a chat application's usable context are separate things to verify. Anthropic lists the app breakdown in its Claude context-window help article (opens in a new tab). Check the section for the product you use; do not treat a developer specification as a promise about every account or integration.
Write down the route before troubleshooting: browser chat, Claude Code, or an application sending model requests. Then note the exact error and whether it appeared during upload, before generation, or after an answer had begun. Those observations help distinguish a file restriction, request-capacity problem, and interrupted response.
| Question | Where to look | What to record |
|---|---|---|
| What can this model process? | The exact Opus model specification. | Model identifier and required input and output. |
| What can this Claude app use? | The app context guide and the selected settings. | Application and model selection. |
| Why must I wait before continuing? | The usage notice and account information. | The message shown and the work left unfinished. |
When comparing two routes, keep the source excerpt and question the same. A vague report that Opus handled a file elsewhere leaves too many variables. A small reproducible example is easier to investigate than an entire project attached without an explanation.
How are context capacity and usage limits different?
Anthropic distinguishes usage limits, which govern interaction over time, from length limits, which concern the material a conversation can handle. Its usage and length explanation (opens in a new tab) also describes how long conversations may be managed. Use the wording of your actual notice to decide which problem you have.
- If the notice concerns input or conversation length, reduce the material needed for the next task or begin a focused session.
- If the notice concerns usage, check the account guidance rather than repeatedly resending a shorter version of the same task.
- If an upload fails, inspect the file requirements before changing the model.
- If the response ends early, check what was actually completed before asking the assistant to continue.
A restart can reduce conversational clutter, but do not assume it solves an account restriction. Equally, waiting does not make an oversized request smaller. Keep the two diagnoses separate so you do not waste time repeatedly trying a remedy for the wrong limit.
How much of your source material belongs in one request?
An AI context window is temporary working space. Tokens, the small pieces of text a model processes, give a more useful measure than document count. Still, the first planning question is about relevance: which sources are needed to answer the current question correctly?
Say you are reviewing a proposal with several appendices. Begin with the decision you need: whether the revised delivery plan meets the approved requirements. Include those requirements and the relevant schedule. Bring in another appendix when the analysis depends on it, while marking any conclusions that still need evidence.
Do not substitute a summary for the original when exact wording matters. A summary can tell the assistant where to look, but the assistant should inspect the relevant source before quoting it or making a consequential comparison. Keep enough information to verify the answer, even when the starting context stays short.
Adjacent guides cover Claude Projects context, general Claude capacity, Claude versus Gemini, Haiku 4.5 context, and Fable context.
What should you save before an Opus session stops?
Save a work checkpoint with four parts: the goal, what changed, what was verified, and what happens next. Add source locations and unresolved questions where they affect the next action. Do not write “review complete” if the assistant only examined the first section of a document.
Separate a person's decision from an assistant's suggestion. If the assistant recommends a different delivery date, the saved record should say that approval is pending. The next session needs to know what it may proceed with and what still requires a person to choose.
Ask the next assistant to summarize that checkpoint before continuing. Check whether it names the unfinished task and its outstanding evidence. The fenbs handoff workflow keeps Where we left off alongside tasks and decisions, so those distinctions remain available after the conversation changes.
Restart from the work you reviewed
Keep the approved brief, active tasks, and source references in fenbs. Have the connected assistant read the project context before it resumes the Opus task, then save progress when it stops. You maintain a reusable starting point without copying every previous exchange into the next window.
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 and add the project that needs continuity between sessions.
- Add the unfinished tasks with their current state and the next check. Record what people approved in Decisions and rules, leaving proposals clearly unresolved.
- Connect a supported AI app over MCP, a standard way for an assistant to use an outside tool. Ask it to read the project context and identify the task it will resume.
- Before stopping, ask the assistant to save task progress and Where we left off. Review source references, completed checks, and anything waiting on a person.
fenbs is free for one person, with no credit card. It does not increase Opus capacity, reset usage allowances, or import Claude conversations. If your work is a single self-contained question, keeping it in one focused chat may be sufficient.
Try the restart with an ordinary task before relying on it for a larger review. The next assistant should know what remains without pretending it has read sources you have not provided. Explore the context engineering guides for more ways to keep that starting material useful.