If I'm reading this right, they ran open source models through each harness (Deepseek, GLM etc) and then compared results. It's my understanding that Claude and Codex are highly optimized for their own models: system prompt, tooling, etc etc. So shoehorning in another model that they weren't designed for feels like setting them up for failure, especially compared to OpenCode (optimized for general use with [mostly?] open source models AND Deepseek's own harness running...it's own model! Doesn't seem like a well-set-up test.