Anthropic has upgraded its Claude Sonnet AI model to version 4.6. The new iteration introduces significantly enhanced computer-use skills, extended reasoning capabilities, and an expanded 1 million token context window, currently in beta. This update brings Sonnet closer to the performance of Anthropic's flagship Opus model.
For users on Free and Pro plans, Sonnet 4.6 is now the default with unchanged pricing. The model demonstrates major improvements in computer-use skills, approaching previous Opus-level performance. This advancement is particularly notable in automating computer user interfaces, with capabilities that allow it to navigate complex spreadsheets, fill multistep web forms, and synthesize information across multiple browser tabs.
Box Inc. CTO Ben Kus noted that Sonnet 4.6 outperformed its predecessor in heavy reasoning Q&A by 15 percentage points in tests involving enterprise documents. This progress places a significant number of human-capable tasks within reach.
The expanded 1 million token context window is a significant boon for developers, enabling the AI to ingest entire codebases and understand full scopes of dependencies at once. This allows for more efficient code modification and consolidation. For business applications, the model can now hold lengthy contracts or numerous research papers in memory, referencing them simultaneously for analysis.
Within the API, Sonnet 4.6 now supports adaptive and extended thinking, along with context compaction, allowing users to optimize for cost-to-performance and continuous execution. Additionally, Claude's web search and fetch capabilities now automatically write and execute code to filter results, with code execution, web fetch, memory, and programmatic tool calling now generally available for production applications.
Claude Cowork, a new MacOS desktop application with a Windows version forthcoming, allows the AI to interact with files on users' computers, controlling the mouse, keyboard, and browser for tasks like file organization and web browsing. Anthropic emphasizes ongoing work to improve resistance to manipulation and hallucination, crucial for AI operating with full computer control.