Kimi K3.1 Pro isn't it
I love the Moonshot team. They still feel like the build-it-in-the-garage culture that does not exist in San Francisco anymore. So I wanted K3.1 Pro to be good.
I gave it a hard task: audit a legacy, layered codebase and turn it into a technical manual I could actually explore. The regular model was rate limited on the API, so I ran the high-speed variant, which is twice the price. It spawned sub-agents that spawned their own sub-agents, hit the 15 concurrent session limit, and burned $14.81. At regular pricing that is about $7.40, and I still did not get a report I would keep.
The output was ugly, the research was thin, and when I gave it $3 to write a plan it ran out of money first.