Anthropic has introduced the latest iteration of its mid-size Sonnet model, continuing its established pattern of updates every four months. In the announcement, the company highlighted enhanced capabilities in coding, following instructions, and computer-based tasks. Sonnet 4.6 will serve as the standard model for users on both the Free and Pro subscription plans.

The beta version of Sonnet 4.6 will feature a context window of 1 million tokens, doubling the previous maximum available for the Sonnet series. Anthropic noted that this expanded capacity is sufficient to process entire code repositories, extensive legal contracts, or multiple research papers within a single query.

This release follows the debut of Opus 4.6 by just two weeks, with an upgraded Haiku model expected in the near future. The launch is accompanied by a series of new benchmark records, such as OS World for computer use and SWE-Bench for software engineering. Particularly notable is its achievement of 60.4% on the ARC-AGI-2 benchmark, which evaluates abilities associated with human-like intelligence.

This performance positions Sonnet 4.6 ahead of many similar models, though it still remains behind others like Opus 4.6, Gemini 3 Deep Think, and a refined variant of GPT 5.2.