Report: US AI Models’ Capability Gains Accelerate, Anthropic’s Doubling

The tension here is internal to the lab, not between rivals: the same organization whose leadership publicly argued for decelerating frontier development is now the cited example of acceleration. That gap between stated caution and measured capability gains is what gives the report its weight, since it suggests competitive pressure is outpacing voluntary restraint even among the most safety-vocal developers. The open question is whether evaluation findings of this kind start feeding into policy or governance debates around frontier model pacing, or remain a technical data point.