You are in the middle of work, and then you hit a wall. The conversation stops, or the responses start to deviate from the document you just sent. Many people conclude one thing: your quota is exhausted.
However, there are two different walls, and the solutions are opposite. Misidentifying the wall means waiting for something that will never reset, or starting a new conversation when the problem is not there.
Usage limits determine how much
Usage limits measure how often you use Claude within a certain timeframe. The amount is influenced by the length and complexity of the conversation, the features used, the model selected, and the effort level set.
One thing that is often overlooked: usage on claude.ai, Claude Code, and Claude Desktop counts towards the same limit. So switching applications does not give you a new quota.
If this is what you hit, starting a new conversation will not help. What helps is waiting for a reset, upgrading your package, or adding usage credits.
Length limits determine how large a single conversation is
Length limits relate to the context window, which is the working memory for a single conversation. It does not only contain your last question.
What is also counted in the context window:
- system instructions and project instructions
- the entire conversation history, not just the last message
- any attachments, documents, and images
- tool definitions and results of tool calls
- the answer that Claude is currently writing
The last point is the most frequently forgotten. A portion of the context window is always reserved for the answer, so the longest possible conversation is always slightly smaller than the window size.
If this is what you hit, waiting will not help at all. What helps is starting a new conversation or moving materials into Projects.
The numbers, and why numbers are not the answer
As of September 2026, official documentation mentions three tiers for conversations in paid packages: 1 million tokens on the latest models like Opus 5 and Sonnet 5, 500 thousand tokens on some previous models, and 200 thousand tokens outside of that. As a rough estimate, 200 thousand tokens is equivalent to about 500 pages of text.
These numbers change every time a new model is released. Treat the list above as a snapshot, not a permanent benchmark, and check the official page when you really need the exact numbers.
Longer is not automatically better
This is the part that is rarely mentioned, yet it is the most important for analytical work.
Official documentation states that as the number of tokens increases, the accuracy and memory of the model decrease. This phenomenon is called context rot. This means cramming the entire archive into one conversation is not a safe strategy just because it fits technically.
The practical consequence: selecting the materials included is as important as how much space is available. Five relevant pages outweigh two hundred pages that are mostly unused.
What really saves space
Official documentation mentions several ways, all about reducing the unnecessary:
- Use Projects. Projects utilize retrieval augmented generation, so only relevant parts are loaded into the context window.
- Simplify project instructions. Fill it with general context, key guidelines, and Claude's role, then save specific task instructions in its own conversation.
- Clean up project files that are no longer in use.
- Turn off tools and connectors that are not currently needed. Both are wasteful of tokens.
- Lower the effort level or turn off extended thinking for routine tasks.
There is also automatic context management available, which summarizes early messages when the conversation approaches the limit, while keeping the complete history stored. This feature requires active code execution.
Quick ways to identify which wall
If the problem arises in any conversation you open, including new and short ones, it is the usage limit. If a new conversation flows smoothly while the old one is stuck, it is the length limit.
Sources
- Context windows, Claude Platform Docs, Anthropic
- How do usage and length limits work?, Claude Help Center, Anthropic
- How large is the context window on paid Claude plans?, Claude Help Center, Anthropic