85% cost reduction is believable, but I'd push back a bit on attributing it mainly to the graph engineering vs. Kimi K3's pricing itself — K2 was already priced aggressively below GPT-4 class models, so some of that delta might be model selection, not architecture. Worth isolating the two variables if you haven't: run the same graph setup on a comparably priced model and see how much of the accuracy gain survives.