Claude Sonnet 3.7 has demonstrated improved performance with its extended thinking mode, recently climbing to second place in testing rankings. Tests were conducted using different token limits, including 20K and 40K, with no discernible difference in results. Performance metrics indicate that the base model achieved a success rate of 12.2% at a cost of $0.05 per task, while the extended thinking modes showed varied results: 11.6% at 1K tokens ($0.07/task), 21% at 8K tokens ($0.21/task), and 28.6% at 16K tokens ($0.33/task). The performance of Claude Sonnet 3.7 is reported to be comparable to the o3-mini model, albeit with a slightly higher cost per task. Additionally, the extended thinking feature will be included in the next update of the Live AI Assistant, which will provide insights into its thinking process. The extended thinking mode supports an output limit of 128K tokens, although longer thinking runs may require several minutes to complete.