Anthropic releases Sonnet 4.6 with 1 million token context window

Anthropic has released a new version of its midsized artificial intelligence model, Sonnet 4.6, keeping pace with the company’s four-month update cycle. In a post announcing the new model, Anthropic emphasized improvements in coding, instruction-following, and computer use. Moving forward, Sonnet 4.6 will serve as the default model for users on both the Free and Pro plans, replacing the previous version of the model for these users.

The beta release of Sonnet 4.6 introduces a context window of 1 million tokens, which is twice the size of the largest window previously available for the Sonnet model. In artificial intelligence, a context window represents the amount of data a model can process at once. This data is measured in tokens, which are the basic units of text processed by AI models. To illustrate the scale of this expanded capacity, Anthropic described the 1 million token context window as “enough to hold entire codebases, lengthy contracts, or dozens of research papers in a single request.” This expansion allows users to input significantly larger datasets in a single query.

The launch of Sonnet 4.6 comes with a new set of record benchmark scores across several industry-standard evaluations used to measure artificial intelligence performance. These evaluations include the OS World benchmark, which measures performance in computer use, and the SWE-Bench evaluation, which measures software engineering capabilities. Most notably, Sonnet 4.6 achieved a score of 60.4% on the ARC-AGI-2 benchmark, an evaluation meant to measure skills specific to human intelligence. The model’s performance on these tests highlights its comparative standing in the market, placing it above many of its peers:

• The 60.4% score achieved on the ARC-AGI-2 benchmark places Sonnet 4.6 ahead of most comparable models.
• Despite these gains, Sonnet 4.6 still trails models like Opus 4.6, Gemini 3 Deep Think, and one refined version of GPT 5.2.

The rollout of Sonnet 4.6 comes just two weeks after the launch of Opus 4.6, continuing a rapid sequence of model updates. Following these consecutive releases, an updated Haiku model is likely to follow in the coming weeks ahead.

Why it matters

Anthropic’s rapid four-month release cycle and the expansion of context windows to 1 million tokens signal an aggressive push to maintain competitiveness against models from Google and OpenAI.