Two numbers on this model that don't sit comfortably with each other. The Pro config posts coding scores near the top of every board, 80.6 on SWE-bench Verified and 93.5 on LiveCodeBench. Then CAISI ran it across a spread of domains and landed on it being roughly eight months behind the US frontier, around where GPT-5 was. DeepSeek's own framing at launch put it two months back, right behind the frontier at the time. Same weights, very different verdicts. The way I read it, both are right and ...