The gap between the best AI models built in the United States and those built in China has collapsed to a record-low 6% on capability benchmarks, according to a new Bloomberg Intelligence analysis reported by The Next Web this week, a threshold that policymakers in Washington had, until recently, treated as a comfortable multi-year buffer. The finding, which tracks the top AI models from both countries against a standardized evaluation suite, marks the culmination of an 18-month sprint in which Chinese laboratories have shed what was once a lead measured in years and now measure it in weeks.
The number itself is a composite: a rolling average across reasoning, coding, math, and multimodal benchmarks that pit US frontier systems OpenAI’s Fable 5, Anthropic’s Mythos 5, Google DeepMind’s Gemini 3.1 Pro against the strongest open- and closed-weights systems from Chinese developers. As late as late 2024, that same composite showed a roughly 40% gap. A year ago it was 20%. Today it is 6%, and on several sub-benchmarks the ranking has already flipped.
From Years to Weeks
The story of how the gap closed is, at this point, a familiar one, but the pace has stunned even veteran analysts. When ChatGPT launched in November 2022, US labs were widely judged to be more than three years ahead of anything China could ship. DeepSeek’s R1 release in early 2025 compressed that lead to a matter of months. The one-two punch of Alibaba’s Qwen 3.8 preview and Moonshot’s Kimi K3 in July 2026 has now reduced it to weeks, according to Ryan Fedasiuk of the American Enterprise Institute, quoted in a related South China Morning Post analysis. “I don’t think anybody truly knows what’s going on,” Fedasiuk said of the speed at which Chinese labs are iterating.
Two releases in particular are driving the newest revision. Kimi K3, unveiled earlier this month by Moonshot AI, is a 2.8-trillion-parameter mixture-of-experts model whose coding and agentic performance triggered such heavy demand that Moonshot was forced to suspend new subscriptions due to a compute shortage. Former White House AI advisor David Sacks noted publicly that “Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of ‘cyber guardrails.'” The company’s annual recurring revenue crossed $300 million within weeks of launch, and it is now preparing a Hong Kong IPO at a $50 billion valuation.
The second is DeepSeek V4, whose flagship V4-Pro-Max variant scored 80.6% on SWE-bench Verified — the highest ever recorded for an open-weights model, tied with Gemini 3.1 Pro. V4-Pro pairs 1.6 trillion parameters with 49 billion active parameters and a one-million-token context window, priced at $0.435 per million input tokens. Its smaller sibling, V4-Flash, is roughly one-hundredth the price of OpenAI’s GPT-5.6 Sol at the output tier. Legacy DeepSeek endpoints were retired on July 24, forcing every enterprise customer to migrate to the new family in a single day. Coming from a lab that Liang Wenfeng bootstrapped with a reported $3 billion of his own capital, V4 is a decisive statement that China’s open-weights ecosystem is no longer trailing — it is setting the price and, increasingly, the performance ceiling.
Open Weights, Global Reach
What makes the 6% figure politically potent is that most of the Chinese gains are landing in the open. Qwen surpassed one billion cumulative downloads on Hugging Face by March 2026. Andreessen Horowitz estimates that roughly 80% of US AI startups now build on Chinese base models at some layer of their stack. As we reported in our coverage of the recent Washington probe into US firms using Chinese models, that dependency has become a live national security question.
Ben Hayum of the Center for a New American Security told SCMP that most Western analysts had assumed Beijing would clamp down on open-source releases once frontier models crossed obvious security thresholds. “They haven’t,” he said. Instead, Chinese labs have doubled down, a strategy underscored last week at the World Artificial Intelligence Conference in Shanghai, where Xi Jinping delivered a keynote pitching an “inclusive” and “AI for all” vision that drew signatures from 28 Global South countries.
Washington’s Response Fractures
The tightening race has splintered US policy. The Trump administration forced Anthropic to suspend its Fable 5 and Mythos 5 releases last month over national security concerns, a move that critics argue is now backfiring. If US labs are throttled at home while Chinese labs ship freely abroad, the argument goes, Washington has effectively ceded the standard-setting layer of the global AI stack.
Axios reported that Kimi K3’s launch emboldened parts of the administration to push for a sweeping ban on Chinese open-weight models running on US infrastructure. Others, including the more than 200 Silicon Valley founders who signed an open letter earlier this week, argue that such a ban would be self-defeating. Treasury Secretary Scott Bessent has tried to occupy a middle position, publicly supporting open-source while threatening sanctions on “adversarial distillation” and hinting that US firms may soon be required to disclose their use of Chinese models.
Complicating matters further, the Center for AI Standards and Innovation, the NIST unit tasked with evaluating Chinese models on issues such as censorship and “alignment with Communist Party talking points”, quietly stopped publishing those findings, and director Chris Fall departed after only three months on the job.
Beyond Capability
Fedasiuk argues that Washington needs to move past a benchmark-by-benchmark contest, it may not be able to win outright. “The competition is going to shift to infrastructure buildout,” he said, pointing to US deployments in the UAE, Saudi Arabia, and other markets where American models still enjoy default-vendor status. That framing dovetails with an internal administration argument that the September summit between Presidents Trump and Xi should focus less on freezing model capability and more on securing compute, energy, and diffusion advantages that are harder to reverse-engineer.
Wang Yaqiu of the University of Pennsylvania’s Project on the Future of US-China Relations predicts that Washington’s rhetoric will pivot in the coming months from capability comparisons to ideological ones, highlighting censorship, alignment, and provenance rather than raw performance. That may prove a more durable message. But as the 6% number circulates, it is becoming harder to argue that the underlying capability gap will do the political work for them.
Whether measured in weeks, benchmark points, or Hugging Face downloads, the frontier has narrowed to a photo finish — and China’s runners are still accelerating.