I redid Dan's zstd rust rewrite benchmark for Anthropic models and was shocked to find haiku 5.5 can do it. The latest models are nuts.

https://taktile.com/blog/property-based-testing-claude

Or the code has been added to the training data.

For me the most recent Claude models are consistently delivering complex PRs that only receive minor and nit level feedback from multiple automated AI reviewers. I rarely plan anymore, but still submit fairly long dictated prompts. It took me a couple of weeks to realize this was the case, and I haven't done any analysis on pre-5.5 PRs, but it definitely feels like they quietly crossed some kind of new threshold of quality and accuracy in coding.