Bloomberg Intelligence puts the US–China frontier-model gap at about 3% after DeepSeek V4.1 Flash

Bloomberg Intelligence analysts estimate the performance gap between the top US and Chinese models has narrowed to roughly 3%, from about 15% earlier this year, driven largely by DeepSeek's V4.1 Flash, which ranks sixth globally on LiveBench. The same analysis notes the economics remain brutal: most of China's 1,100-plus model vendors are unprofitable, ByteDance's Doubao is reportedly the only profitable consumer assistant, and DeepSeek and Tencent give their chatbots away. Near-parity open models at a fraction of the price keep downward pressure on US API pricing and strengthen the case for open-weight stacks in cost-sensitive products.

Companies mentioned

DeepSeek