Anthropic has unveiled Claude Sonnet 4.6, a major upgrade designed to bridge the gap between mid-tier efficiency and top-tier reasoning. This iteration is now the default for Claude.ai users, offering a massive one-million-token context window—the amount of information the model can "keep in mind" at once—which is sufficient to process entire codebases or dozens of research papers in a single prompt.

A standout feature is the model’s refined "computer use" capability. Rather than relying on rigid backend connectors, Sonnet 4.6 interacts with computers much like a human does, by visually interpreting screens and simulating mouse clicks or keystrokes. In verified benchmarks, it demonstrated human-level proficiency in navigating complex spreadsheets and multi-step web forms, marking a significant step toward truly autonomous AI agents capable of handling messy, real-world office workflows.

For developers, the model introduces "context compaction," a feature that automatically summarizes older parts of a conversation to prevent hitting memory limits. It also supports "adaptive thinking," allowing the model to adjust its internal reasoning effort based on task difficulty. Early testers report that the model is notably less prone to "laziness" and hallucinations, frequently outperforming Anthropic’s own previous flagship, Opus 4.5, in complex coding tasks.

The update also integrates the Model Context Protocol into Excel, enabling the AI to pull data directly from external financial sources without leaving the spreadsheet. Despite these gains in autonomy, Anthropic emphasizes that safety remains a priority, with improved defenses against prompt injection attacks—where malicious instructions are hidden within websites to hijack the model's behavior.

Anthropic has unveiled Claude Sonnet 4.6, a major upgrade designed to bridge the gap between mid-tier efficiency and top-tier reasoning. This iteration is now the default for Claude.ai users, offering a massive one-million-token context window—the amount of information the model can "keep in mind" at once—which is sufficient to process entire codebases or dozens of research papers in a single prompt.

A standout feature is the model’s refined "computer use" capability. Rather than relying on rigid backend connectors, Sonnet 4.6 interacts with computers much like a human does, by visually interpreting screens and simulating mouse clicks or keystrokes. In verified benchmarks, it demonstrated human-level proficiency in navigating complex spreadsheets and multi-step web forms, marking a significant step toward truly autonomous AI agents capable of handling messy, real-world office workflows.

For developers, the model introduces "context compaction," a feature that automatically summarizes older parts of a conversation to prevent hitting memory limits. It also supports "adaptive thinking," allowing the model to adjust its internal reasoning effort based on task difficulty. Early testers report that the model is notably less prone to "laziness" and hallucinations, frequently outperforming Anthropic’s own previous flagship, Opus 4.5, in complex coding tasks.

The update also integrates the Model Context Protocol into Excel, enabling the AI to pull data directly from external financial sources without leaving the spreadsheet. Despite these gains in autonomy, Anthropic emphasizes that safety remains a priority, with improved defenses against prompt injection attacks—where malicious instructions are hidden within websites to hijack the model's behavior.