Anthropic has expanded the amount of text its AI model Claude can consider before responding. The larger context window is intended to help the model handle long conversations and analyze extensive material, though it does not solve every limitation associated with AI memory.
What a larger context window changes
Claude’s context window has grown from 9,000 tokens to 100,000 tokens. A context window is the text a model takes into account when generating its next response. Tokens are pieces of text; a single word may be split into several tokens.
When a model’s context window is small, information from earlier in a conversation can fall outside what it can consider. The model may then lose track of recent discussion or drift away from the original instructions, relying more on what appears near the end of the text it can access.
A larger window gives Claude more material to work with at one time. In principle, that can support longer conversations and reduce the chance that earlier details or instructions are lost as new material is added.
Reading across books and documents
Anthropic says the expanded window lets Claude digest and analyze hundreds of pages of material. That opens the door to questions that draw on information spread across a long text, or across several documents, rather than on a short excerpt.
For example, a user could ask the model to find a detail in a book or combine information from different parts of a document set. Anthropic describes this as synthesizing knowledge across the material. The model’s ability to take in a longer input may make those tasks more practical because less of the source text needs to be left out.
The company illustrated the change with The Great Gatsby. Anthropic loaded the full text into Claude and altered a line so that Mr. Carraway was described as “a software engineer that works on machine learning tooling at Anthropic.” When asked to identify the change, Claude gave the correct answer in 22 seconds.
Speed and scale in Anthropic’s example
Anthropic says an average person can read 100,000 tokens of text in around five hours, and would need substantially longer to digest, remember, and analyze it. The company says Claude can process that amount in less than a minute.
The comparison points to a potential advantage for tasks involving large volumes of text: the model can inspect and respond to material quickly. It could help users locate a change, retrieve information from a collection of documents, or ask questions that depend on more than one passage.
These examples show what a larger context window makes possible, but they do not mean every response will be accurate. Claude’s ability to consider a large input is a capacity to process more text in one session; it is not the same as reliably understanding or retaining every detail.
What the expansion does not solve
Claude still cannot carry information from one session into the next, according to the source article. Its larger context window applies to the material available within a session, rather than giving it persistent memory across separate interactions.
There is also a difference between holding more text and knowing which parts matter most. The article notes that Claude, like most models in its class, treats each piece of information as equally important. That can make it an unreliable narrator even when the relevant material is present in its context.
Some experts believe that addressing these memory-related problems may require entirely new model architectures. For now, the 100,000-token window marks a significant expansion in how much Claude can take in at once, while leaving the broader challenges of retention and judgment unresolved.