On May 11, 2023, Anthropic announced that it had expanded the context window of its Claude language model from 9,000 tokens to 100,000 tokens, roughly equivalent to 75,000 words. The change allows businesses to submit hundreds of pages of material for Claude to analyze, and enables conversations with the model to continue for hours or even days.
Fast ingestion and retrieval
Anthropic notes that an average person would take about five or more hours to read 100,000 tokens of text and then considerably longer to digest and analyze it, while Claude can process that volume in seconds. In one demonstration, the entire text of The Great Gatsby (72,000 tokens) was loaded into Claude-Instant and a single line was altered to state that Mr. Carraway was "a software engineer that works on machine learning tooling at Anthropic." When asked to spot the change, the model returned the correct answer in 22 seconds.
Practical applications of a larger context window
Anthropic outlines several practical uses enabled by the 100K token context window:
- Digesting, summarizing, and explaining dense documents such as financial statements or research papers
- Analyzing strategic risks and opportunities for a company using its annual reports
- Assessing the pros and cons of proposed legislation
- Identifying risks, themes, and argumentative patterns across legal documents
- Reviewing hundreds of pages of developer documentation and surfacing answers to technical questions
- Rapidly prototyping by loading an entire codebase into the context and intelligently modifying or building on it
The company says that for complex queries this approach can outperform vector-search-based methods, because Claude can synthesize information across many parts of the input and follow user instructions like a human assistant.
Audio and partner demonstrations
Anthropic's partners have highlighted the business implications of larger context windows. For example, 100,000 tokens corresponds to roughly six hours of audio — AssemblyAI demonstrated this by transcribing a long podcast into 58,000 words and using Claude for summarization and Q&A.
Availability
The 100K token context windows are available now through Anthropic's API. The announcement points users to documentation on which model versions to call, and provides a way to request access for those not yet using Claude.
Conclusion
Expanding the context window substantially increases the model's ability to process large documents and synthesize information from many sources. The provided examples and partner demos indicate practical applications across business, legal, research, and software-development workflows.
