February 18th.Anthropic It's on a scale, Sonnet ModelThe new version continued the company ' s four-month update. In the announcement of the new model, Anthropic highlighted its improvements in code capabilities, command compliance and computer operations。
Sonet 4.6 will be the default model for free and Pro users。
The Sonnet 4.6 test version will support 1 million token , twice as much as the largest window in the previous Sonnnet. Anthropic indicates that the new context window "single requests can accommodate a complete code library, long contracts or dozens of research papers"。
Only two weeks before the launch of Opus 4.6, the new Haiku model is expected to follow up in the coming weeks。
1AI noted that the model was high in a number of baseline tests, including OS World, which was used to assess computer performance, and SWE-Bench, which was used for software engineering. And the brightest result is that it reaches 60.4% in the ARC-AGI-2 test for measuring human intelligence. This score allows Sonet 4.6 to go beyond most of the same-grade models, but still lags behind flagship models of Opus 4.6, Gemini 3 Deep Think and a fine-tuned version of GPT 5.2。