China’s DeepSeek previews new AI model a year after jolting US rivals

The one-year framing matters more than the model itself. DeepSeek's first disruption forced US labs to justify their infrastructure spending; V4 arrives at a moment when that spending has only accelerated, making the cost-efficiency argument DeepSeek represents structurally harder for incumbents to dismiss a second time.

This lands directly on top of coverage from mid-April showing US labs responding to pressure through acquisition and consolidation rather than model releases. The TechCrunch pieces from April 17 on OpenAI's shopping spree documented a company pivoting toward owning distribution and consumer apps, not racing on raw model capability. If DeepSeek V4 genuinely closes the gap on coding benchmarks, that strategic bet looks more defensible in hindsight: compete on integration and reach, not on model scores alone. Anthropic's simultaneous push into government cybersecurity work (Claude Mythos Preview, covered April 17) suggests a similar instinct to find protected verticals rather than fight on open benchmarks.

Watch whether any of the three named US labs responds with a capability release or a benchmark rebuttal within 60 days. Silence would suggest they've accepted the open-source parity framing and are competing elsewhere; a direct counter-release would signal the benchmark race is still the primary front.