We do spec based development so my eval for all new models is one of the specced products.

We do spec based development so my eval for all new models is one of the specced products. Haiku 5.5 (at max) surprisingly did a great job following the spec-based approach and developing an almost bug-free game (backend/fe/pg-db/feed/redis/...). The game spec path is around 2,784 lines. As I was watching, I saw it's related to its good browse-use along with better coding abilities. I emphasize browse-use as that's how these models find their bugs and fix them. Older and similar models (luna, older sonnets/haikus, ...) never reached this level of low-bug/complete product.