Anthropic 發布 Sonnet 4.6

• Context Window: Supports 200K tokens standard with 1M token context window available in beta; context compaction feature automatically summarizes older context during long conversations • Output Capacity: 64K maximum output tokens for structured responses • Thinking Capabilities: Supports both extended thinking (deliberative reasoning) and adaptive thinking (contextual reasoning adjustment) • Effort Parameter: Introduces effort levels (low, medium, high, max) allowing developers to balance speed, cost, and performance • Web Tools: Web search and fetch tools now automatically write and execute code to filter and process results, improving token efficiency • Tool Availability: Code execution, memory, programmatic tool calling, tool search, and tool use examples now generally available on API • Safety Architecture: Demonstrates improved resistance to prompt injections with behavioral audits showing emotional stability metrics • Coding Improvements: Enhanced consistency, instruction following, and code review capabilities; developers with early access prefer it over Opus 4.5 from November 2025

Sonnet 4.6's performance parity with Opus-class models on economically valuable office tasks suggests a flattening of capability tiers, potentially disrupting premium pricing models. The 72.5 OSWorld benchmark score represents a 2.6x improvement over Sonnet 3.7, indicating accelerating progress in agentic computer use—a critical capability for autonomous task automation. Expanded free-tier access with advanced features (file creation, connectors, compaction) may drive broader adoption and developer ecosystem growth. The emphasis on safety improvements and prompt injection resistance addresses enterprise deployment concerns. Anthropic's consistent four-month release cadence and rapid capability improvements position the company to maintain competitive pressure against OpenAI and other AI providers in the mid-market segment, where cost-performance tradeoffs are critical.