Evals - anyone have Claude results?

[leppikallio]

Weird question, I know and the reason for the question (or rather why I am not running them myself to figure out) is longer story involving bitter feelings and salt poured in countles tiny paper cuts and... anyway, there's a reason to come up with the topic.

I wired the Gas Town with OpenCode & GPT models / model variants out of curiosity, full well knowing that GPT most probably won't behave nicely. So then I ran the evals and results do not exactly encourage. The failure modes & variety makes kind of confirms the "not-fit-for-the-purpose".

But, indeed out of curiosity, what's the Claude's track record with the current GT & evals?

(I can of course attach the "analysis" GPT created if someone is curious)