The gap between open and closed AI models has both collapsed and widened. On composite benchmarks, the distance is 6 points. On crowdsourced preference, it is 29 Elo points. The answer to which side is winning depends entirely on which question you ask, and that is the most important finding of the year.

AI systems can now plan, call tools, coordinate with each other, and take actions in the world without a human approving each step. Production adoption is accelerating. Governance is not. This is the state of the shift from assistant to autonomous agent, and what it means for how we work, learn, and stay in control.

AI systems now score above 90 percent on medical licensing exams, bar exams, and university assessments. But when the same models face real-world clinical tasks, their scores collapse to under 45 percent. The tests were designed for humans, not machines, and the credentialing infrastructure that protects public safety is becoming unreliable at the exact moment AI capability is accelerating.