Anthropic has launched Claude Sonnet 4.6, a significantly improved AI model with a million-token context window, enhanced coding, tool use, and agentic abilities, now available as the default on their free plan. It excels at real-world knowledge work tasks, outperforms previous models and competitors in benchmarks, and features stronger safety measures against prompt injection attacks.
Anthropic has released Claude Sonnet 4.6, positioning it as their new workhorse AI model with significant improvements over Sonnet 4.5. The model now features a million-token context window, enhanced coding abilities, better tool use, and improved agentic capabilities. Notably, Sonnet 4.6 is now the default model on Anthropic’s free plan, with pricing unchanged from the previous version. The model is designed for real-world tasks, excelling at activities like creating PowerPoint presentations and manipulating Excel files within the Claude Co-Work environment, making it especially powerful and fast for knowledge work.
Benchmark results show that Sonnet 4.6 delivers substantial performance gains across various tasks compared to its predecessor. For example, its OS World benchmark score jumped from 61.4% to 72.5%, reflecting its improved ability to interact with computer environments in a human-like manner—using virtual mouse clicks and keyboard inputs rather than specialized APIs. The model also demonstrates significant improvements in agentic terminal coding, computer use, tool use, and financial analysis, often outperforming or matching Anthropic’s higher-tier Opus 4.6 model and other leading AI models like Gemini 3 Pro and GPT-5.2.
Safety and security remain a focus for Anthropic, especially regarding prompt injection attacks, where malicious actors attempt to manipulate the AI by embedding harmful instructions in text. Anthropic claims that Sonnet 4.6 is more resistant to such attacks, with safety evaluations showing major improvements over Sonnet 4.5 and performance on par with Opus 4.6. The model is deployed under AI Safety Level 3 (ASL 3), indicating it poses a higher risk of catastrophic misuse compared to non-AI baselines, but does not yet reach the thresholds for fully automating research or assisting in the development of high-consequence weapons.
Sonnet 4.6 is particularly optimized for knowledge workers, as evidenced by its top scores in benchmarks related to office tasks, financial analysis, and real-world productivity. In the Vending Bench simulation, where the AI autonomously manages a vending machine business, Sonnet 4.6 dramatically outperformed Sonnet 4.5 by adapting its strategy for greater profitability. Additional product updates include context compaction in beta, improved web search and code execution tools, and enhanced integration with Excel through MCP connectors.
Overall, Sonnet 4.6 blurs the line between Anthropic’s Sonnet and Opus models, with many speculating that it may be based on next-generation training runs. The model’s capabilities make it a standout choice for automating and enhancing knowledge work, though its increasing power also raises new challenges in measuring and ensuring safety. Anthropic encourages users to explore Sonnet 4.6, highlighting its value for entrepreneurs, content creators, and anyone seeking to leverage AI for real-world productivity.