Nice article, mirrors my experience. Last year (around when multimodal 3.5 Sonnet launched), I had run a sizeable number of PDFs through it. Accuracy was remarkably high (99%-ish), whereas GPT was just unusable for this purpose.