This is not a deep dive. This is an AI-generated mix of bullet points and tables with no statistical meaning, similar to the other two blog posts on this website.

As a sibling commenter mentioned, the article seems to be more of an AI-generated outline.

That said, it mostly mirrors my (very subjective) experience, and the observations in it also mirror my experience with Gemini 2.5 Pro. Overall, Sonnet (whether 3.7 or 4) is much better at understanding exactly what my incomplete prompt means and doing what I want it to do.

Sonnet 4 is also a meaningful improvement with Claude Code. Faster, and gets good results quicker, than 3.7.

Overall, I'm currently primarily using Sonnet/Opus 4 for anything complicated, while using Gemini 2.5 Flash for mechanical refactors of many files that I can't easily and quickly do in a mechanical way (like regex search+replace), because it's just so ridiculously fast.

More data would be nice.