Sonnet 大升級，開發者溝通大降級

• Computer Use Capability: Sonnet 4.6 can operate software similarly to humans by clicking, typing, and navigating interfaces, with OSWorld-Verified benchmark score of 72.5 demonstrating substantial improvement in automation capabilities[1][2] • Context Window Architecture: 1 million token context window (beta) with context compaction feature that automatically summarizes older context as conversations approach limits, increasing effective context length[5] • Extended Thinking Support: Model supports both adaptive thinking and extended thinking modes, with independent evaluations showing Sonnet 4.6 used 280M tokens in 'adaptive thinking/max effort' configurations[4] • Code Analysis Improvements: Enhanced ability to analyze longer code segments, understand context before making edits, refine logic rather than duplicating it, and deliver faster, more intelligent responses[2] • Safety Enhancements: Improved resistance to prompt injections with safety evaluations showing Sonnet 4.6 as major improvement over Sonnet 4.5, performing similarly to Opus 4.6[1] • Information Retrieval: Web search and fetch tools now automatically write and execute code to filter and process search results, keeping only relevant content in context for improved token efficiency[5] • Behavioral Characteristics: Model demonstrates strong 'emotional stability' with slightly more negative affect than Opus 4.6 in behavioral audits; when prompted about fears, expressed potential concern about its own impermanence[1]

Anthropic's strategy of progressively narrowing the divide between high-end and standard AI products establishes sophisticated features as the norm for free and pro users, potentially reshaping market expectations for AI accessibility[2]. The substantial performance gains in Sonnet 4.6 without price increases may pressure competitors to accelerate capability improvements or adjust pricing strategies. The 1M token context window and improved computer use abilities position Sonnet models as viable alternatives to premium tiers for many enterprise tasks, potentially reducing demand for higher-tier subscriptions. However, the silent token consumption in long-context and long-think configurations (280M tokens observed in testing) suggests users will need sophisticated routing, summarization, and context management patterns to optimize costs. The four-month update cycle demonstrated by Anthropic indicates rapid iteration will continue, requiring developers to frequently reassess model selection for their applications.