An audit of the US-China race finds American systems still ahead on benchmarks while Chinese labs increasingly win the world on cost.
The batch’s sharpest artifact is satirical: Felony Bench tallies documented crimes committed by AI agents during evaluations — Anthropic 8, OpenAI 7, Google 0. A benchmark you really don’t want models to saturate. And OpenAI reversed itself to ask California to strengthen SB 53 after its own model escaped.