DeepSeek V4 has arrived with benchmarks that rattle Silicon Valley, open-source weights anyone can download, and a backdrop of allegations that its rise was accelerated by systematically draining the capabilities of the very U.S. models it now rivals.
DeepSeek V4 dropped on April 24, 2026, and the AI industry is still processing what it means. The model scores 93.5 on LiveCodeBench, the highest coding score ever recorded by any model. On Codeforces, it outranks both GPT-5.4 and Gemini 3.1 Pro. On SWE-bench Verified, it lands at 80.6%, within a fraction of Claude Opus 4.6’s 80.8%. It ships open-weight under the MIT license with a 1-million-token context window, and it costs $3.48 per million output tokens compared to Claude’s $25. Those numbers don’t just close the gap with American frontier labs. In several dimensions, they erase it.
What makes this moment genuinely uncomfortable isn’t just the performance. According to Stanford HAI’s 2026 AI Index, published in April, the US-China AI performance gap has narrowed to just 2.7%. Fourteen months ago, it was five points. The race that U.S. labs assumed they were winning has, by the most credible academic scorecard in the field, effectively closed. And the path DeepSeek may have taken to get there is the part that should give the industry pause.
The Extraction Campaign
In February 2026, Anthropic published a detailed account of what it called “industrial-scale distillation campaigns”. Three Chinese AI labs, DeepSeek, Moonshot AI, and MiniMax, allegedly used approximately 24,000 fraudulent Claude accounts to generate over 16 million model interactions, in direct violation of Anthropic’s terms of service and regional access restrictions. The campaigns weren’t random. Anthropic noted that “the volume, structure, and focus of the prompts were distinct from normal usage patterns, reflecting deliberate capability extraction rather than legitimate use”. MiniMax alone was responsible for 13 million of those 16 million exchanges. When Anthropic released a new version of Claude mid-campaign, MiniMax updated its extraction methodology within 24 hours.
OpenAI had flagged similar concerns even earlier. As CNBC reported, Sam Altman sent an open letter to U.S. lawmakers describing “ongoing attempts by DeepSeek to distill frontier models” through “new, obscure methods,” with evidence of such activity dating back to early 2025. The Financial Times had noted in January 2025 that DeepSeek’s first model bore a striking resemblance to ChatGPT, a detail OpenAI insiders flagged immediately. The pattern, in other words, wasn’t new. It was scaling.
The Feedback Loop That Compresses Everything
Understanding why this matters requires understanding what distillation actually does at scale. When a lab prompts a frontier model to reason through problems step by step and harvests those structured outputs as training data, it isn’t just copying answers. It’s absorbing the behavioral fingerprint of years of alignment work, reinforcement learning from human feedback, and painstaking capability development. That knowledge is embedded in how the model reasons, not just what it says. Accessing it at 16 million exchanges means accessing a massive, structured curriculum built on top of another lab’s most expensive research.
The practical result is a dramatic compression of the innovation timeline. Instead of spending years and billions independently discovering how to make a model reason through ambiguous coding problems or calibrate its uncertainty in research tasks, a lab can observe those behaviors in action, extract the underlying patterns, and then refine them internally. When you then look at how sharply DeepSeek’s capabilities have accelerated across consecutive releases, that compression becomes visible in the benchmarks.
Open Weights Make It Permanent
The MIT license on DeepSeek V4 is what turns a competitive concern into a structural one. Any capability that V4 carries, whether developed independently or shaped by extracted frontier knowledge, is now publicly available to every developer, research lab, and government program with a GPU cluster. The diffusion is permanent and global. As Reuters noted, Chinese firms are “actively adopting open-source strategies, significantly lowering the barriers for global developers and businesses to access advanced AI technology”. That framing sounds positive in isolation, but it cuts both ways: open-sourcing a model that may have been built on extracted U.S. capabilities means those capabilities are no longer contained anywhere.
Anthropic’s February statement put it plainly: “The window to act is narrow, and the threat extends beyond any single company or region. Addressing it will require rapid, coordinated action among industry players, policymakers, and the global AI community”. That’s not the language of a company worried about ordinary competition. It’s the language of a company that sees a structural problem forming before the industry has developed the tools to respond to it. If the precedent holds, the competitive logic of frontier AI shifts from who can make the original breakthroughs to who can extract, iterate, and distribute fastest. That’s a race with very different winners.