{"slug": "gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years", "title": "Gap Between US And Open Chinese Models Is Now Among The Largest In Last 2 Years", "summary": "The gap between the top US and Chinese AI models has widened to roughly 10 points on Epoch AI's Capabilities Index, the largest since the index began in 2023, with GPT-6 Astra at 168.9 and Moonshot AI's Kimi K3 at 158.7. Epoch AI's data shows every new high-water mark on the composite benchmark since 2023 has come from a US model, and the second-half 2026 releases from OpenAI, Google and Anthropic reversed a two-year convergence that had briefly narrowed the gap to single digits. The index opened in March 2023 with GPT-4 near 110 against Baichuan2-13B near 90, a 20-point gap that DeepSeek's V2, V3 and R1 models and Alibaba's Qwen line closed before the recent US pullback.", "body_md": "Chinese AI models had given quite a scare to US labs — and even the government — but the US is pulling ahead again.\n\nThe distance between the best American and Chinese AI models has stretched back out to levels not seen in roughly two years, according to new data from Epoch AI.\n\nEpoch’s Capabilities Index (ECI), a composite score built from dozens of benchmarks to allow for apples-to-apples comparisons across models and time, shows the top US model now sitting at 168.9, held by GPT-6 Astra. The best Chinese model, Kimi K3 from Moonshot AI, trails at 158.7. That puts roughly 10 points between the two frontiers, the widest the gap has looked since the early days of the index in 2023.\n\nThe chart tells a story of convergence followed by a pullback. Back in March 2023, GPT-4 opened at an ECI of around 110 while China’s best model at the time, Baichuan2-13B, sat near 90, a 20-point gap. Over the following two years, Chinese labs closed in fast. DeepSeek’s V2 and V3 releases, followed by the R1 reasoning model in early 2025, narrowed the distance to single digits at several points, and Alibaba’s Qwen line kept pace step for step with OpenAI’s o1 and o3 models. By some accounts the race had tightened to the point of being nearly a dead heat, with one benchmark comparison earlier this year putting GPT-5.4 Pro, Gemini 3.1 Pro and Qwen3.5-Omni within a few points of each other.\n\nThat convergence appears to have reversed in the back half of 2026. Google’s [rapid climb with Gemini 4 Carbon, which recently topped the Artificial Analysis Intelligence Index to become the top AI model in the world](https://officechai.com/ai/googles-gemini-4-carbon-tops-artificial-analysis-intelligence-index-becomes-top-ai-model-in-the-world/), pushed the US frontier higher. Prior to that, Claude had relased Opus 5.5, and OpenAI had come up with GPT-5.5 Pro and then GPT-6 Astra. On the Chinese side, Qwen3.7-Max and Kimi K3 have kept climbing too, but not fast enough to keep the gap from widening again.\n\nOne detail stands out in Epoch’s dataset: every single step up in the overall global frontier since 2023 has come from a US model. China’s labs have repeatedly caught up to a previous US milestone, sometimes within months, but none of DeepSeek, Alibaba or Moonshot has ever been the one to set a new high-water mark on the index. The leapfrogging has run entirely in one direction.\n\nThis is a notable shift in a race that was, as recently as earlier this year, being described as effectively tied at the top. It’s also a reminder that the “China caught up” narrative that followed DeepSeek-R1’s debut was really a story about one moment in the cycle, not a permanent state. US labs have since pulled further ahead on the back of a wave of releases from OpenAI, Google and Anthropic in the second half of 2026, even as [Anthropic’s own S-1 filing ahead of its Nasdaq listing showed just how dependent its revenue still is on American customers](https://officechai.com/ai/72-of-anthropics-revenue-is-from-american-customers-reveals-s1/).\n\nWhether the gap narrows again will likely hinge on the next DeepSeek- or Kimi-style release. China’s open model ecosystem has shown it can move in sharp bursts rather than steady increments — DeepSeek-R1 alone reshaped the conversation around frontier AI almost overnight. For now, though, Epoch’s data suggests the two-year trend of closing distance has, for the moment, gone into reverse.", "url": "https://wpnews.pro/news/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years", "canonical_source": "https://officechai.com/ai/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years/", "published_at": "2026-10-30 13:01:48+00:00", "updated_at": "2026-10-30 14:17:12.562096+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "generative-ai"], "entities": ["Epoch AI", "OpenAI", "GPT-6 Astra", "Moonshot AI", "Kimi K3", "DeepSeek", "Alibaba", "Google"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years", "markdown": "https://wpnews.pro/news/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years.md", "text": "https://wpnews.pro/news/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years.txt", "jsonld": "https://wpnews.pro/news/gap-between-us-and-open-chinese-models-is-now-among-the-largest-in-last-2-years.jsonld"}}