So I needed to regenerate some Rust code from my spec files as I had added a new feature. I decided to try out Muse Wonder. I sent it the specification, and it spent 54 seconds on a response, which cost me $0.15 (kind of expensive).
The Rust code did not compile, as it had a lifetime issue in a closure. So I had to re-prompt Muse Wonder, which fixed the bug but drove up the cost another $0.11. The strategy Muse Wonder took was more difficult than what other models like GPT, Claude and Gemini use—this could explain its differing result.
But the fact the code did not compile was a disappointment. The Wonder model was anything but wondrous—I ended up moving to Gemini 4 Argon, which implemented the code in a simpler way and was faster and cheaper. If I were advising Meta about this, I would tell them to train on more Rust compiler output and reject invalid code.