China's AI model performance and pricing trends in 2026

The China Academy of Information and Communications Technology, along with the Ministry of Industry and Information Technology's AI Model and Software-Hardware Evaluation Laboratory, and the China AI Industry Development Alliance, released the '2026 Large Model Token Service Performance Monitoring Report' on October 23. The report highlights that domestic token consumption reached 500 trillion per day in June 2026, nearly four times that of December 2025. Domestic models saw a doubling of average TPS to 78.23 tokens per second in August 2026, compared to 38.25 in 2025, though still below international benchmarks. Meanwhile, domestic pricing rose to 13.46 yuan from 6.8 yuan in 2025, while international models averaged 117.8 yuan. The report also noted increased context lengths and performance variations across platforms. The report further found that TTFT increased to 1.26 seconds in August 2026, up from 0.78 seconds in 2025, but remained lower than international models. Open-source models like DeepSeek-V4-Pro showed improved TPS, reaching twice that of earlier versions. Performance differences across platforms were attributed to varying optimization strategies and cost controls. The report aims to guide users in selecting models and improving service quality.