Based on my experience with other platforms, I’ve found that Claude 3.7 makes excellent use of its extended thinking budget to refine and improve its thought process—sometimes iterating multiple times—to deliver higher-quality final outputs. This has been particularly helpful for my use case, which involves crafting highly complex design documents based on extensive references and detailed instructions.

With 32k of thinking budget, Claude 3.7 was able to thoroughly review the document twice, refining it significantly. Among the various reasoning models I’ve tested, it was the only one capable of producing a result comparable to the output from Perplexity Deep Research (which, as you know, employs an agentic approach). While I wouldn’t claim to fully grasp the potential of this feature, it’s undoubtedly a distinctive aspect of Claude 3.7 that I believe many users would be eager to explore.

Best regards