GPT52 now hits almost PhDlevel performance on o

GPT‑5.2 now hits almost PhD‑level performance on olympiad physics, chem, and bio. On paper, it looks like the perfect grad student. Then you give it open‑ended research tasks and it faceplants to around 25% success as per FrontierScience. So it is more like a very confident undergrad who read the slides but never ran an experiment. The twist is GPT‑5 in a tightly scaffolded wet‑lab loop. When you wrap it in a careful protocol with strong constraints, they see a 79× boost in cloning productivity. So now it’s all about how aggressively you can structure the real world into small, verifiable steps before safety, biology, and basic reproducibility push back. If you are building in bio or automation, your edge is probably: 1. System design and workflow scaffolding 2. Domain‑specific constraints and checks 3. How much messy reality you can hide behind APIs and protocols The model is the intern. Your product is the lab.