Claude 4.5 Haiku Codes Like It’s Late for a meeting with God

Claude 4.5 Haiku just landed and it is smoking hot.

In our independent test derived from open-source Java projects, Claude 4.5 Haiku codes as well or better than its big brother Sonnet--and the full-size models from other labs.

This being Anthropic, of course it's priced at a premium. Haiku 4.5 costs just a few cents less than full GPT-5, and 5x more than GPT-5 mini.

But Haiku 4.5 is wicked fast. Twice as fast as the next-fastest model, its big brother Sonnet 4.5. Which makes it over 3x faster than GPT-5 mini, and almost 6x faster than full GPT-5.

I've been using Haiku 4.5 as my primary coding model in Brokk for the last couple days, and I think I'm sticking with it. Sure, it may have decided some unrelated-to-its-task methods were unnecessary and deleted them, but it's so, so fast that I don't mind correcting it.