Claude Sonnet 4.6 is a mid-tier large language model released by Anthropic in February 2026. Designed to balance speed and high-level reasoning, the model effectively bridges the gap between the Sonnet and flagship Opus tiers. It features a significantly expanded 1 million token context window (initially in preview/beta), allowing it to process massive codebases, long-form legal documents, and extensive research archives in a single prompt.
A central feature of the 4.6 architecture is the introduction of Adaptive Thinking and the Effort parameter. This allows developers to control the model's internal computational expenditure via four settings: low, medium, high, and max. The "Non-reasoning, High Effort" configuration refers to the model operating with the high-effort setting while bypassing the explicit generation of extended thinking (Chain-of-Thought) tokens. This mode is optimized for high-quality, direct outputs in complex tasks where transparency into the model's internal reasoning steps is not required.
The model demonstrates advanced capabilities in computer use and autonomous agent planning. On the OSWorld-Verified benchmark, it achieved a score of 72.5%, nearly matching the performance of the Opus 4.6 flagship. It is particularly proficient in multi-step office tasks, financial analysis, and software engineering, where it uses a feature called Context Compaction to automatically summarize older conversation history, preventing context truncation during long-horizon sessions.
Compared to its predecessors, Claude Sonnet 4.6 offers improved instruction following and a reduction in "laziness" or hallucinations. It is engineered for enterprise-scale agentic workflows, providing frontier-level performance in coding and data synthesis at a significantly lower cost-per-token than earlier top-tier models.