The graph shows the dramatic reduction in performance gap between top American and Chinese artificial intelligence (AI) models in the Arena leaderboard. Arena is an Elo-style benchmark that rates chatbot effectiveness, from mid-2023 up to early 2026.
At that time, the American models were leading by a large margin. For example, GPT-4-0314 was scoring around 1,320 on Arena, while the best Chinese model, ChatGLM-6B, was still under 1,000. Towards the end of the two and a half years’ period, the American leading-edge models like o1-preview and Gemini 3 Pro kept improving at a rather fast pace and reached a peak score of 1,503 in January 2026 by Claude Opus 4.6. During the same timeframe, the Chinese AI models improved much more rapidly. One after another, the latest model versions Qwen-14B-Chat, GLM-4-0520, DeepSeek-R1, Qwen3-235B, Dola-Seed-2.0-Preview were getting significant performance increases which led to a very steep increase in benchmark scores.
In early 2025 the Chinese models practically had reached the level of the previously US performance, and with Dola-Seed-2.0-Preview attaining an Arena score of 1,464 in January 2026, the chart reflects the incredible rate at which China is advancing with AI. Even though the U.S. remained the slight frontrunners (1,503 vs. 1,464) the graph mainly emphasizes how China’s progress in AI is really amazing and almost equal to the leading US models around 2026.
