After Two Years of Price War, Chinese Model Prices Nearly Doubled in 2026

October 23 — The AI Institute of the China Academy of Information and Communications Technology (CAICT), the MIIT Key Laboratory for AI Large Models and Hardware/Software Evaluation, and the China AI Industry Development Alliance today (23 October) jointly released the “Large Model Token Service Performance Monitoring Report (2026).”

The report says that with the rapid iteration of large-model technology, token services have become the key point in moving from technology to applications, and that China’s daily token calls keep climbing: in June 2026, domestic daily token consumption was about 500 trillion, nearly four times the level of December 2025.

More and more MaaS (Model as a Service) platform vendors supply large-model token services through public cloud APIs. Users, however, lack intuitive benchmarks for token service performance and face difficulties in choosing between offerings.

Based on monitoring data from January to August 2026, the report analyses platforms’ call success rates, throughput, time to first token, stability, price and context window, and compares vendors’ own official platforms with four open-source models — DeepSeek, Qwen, GLM and Kimi — running on third-party platforms. It further analyses the key factors affecting performance, summarises core industry observations and sets out next steps, aiming to give users a reference for choosing services and the industry a direction for improving them.

The report finds that in August the average TPS (tokens output per second) of domestic models reached 78.23, doubling the 2025 average of 38.25. But the average TPS of typical foreign models in August was 112.82, up from 51.35 in December 2025 — also more than double, and overall better than the domestic figure.

The report finds that the average TTFT (time to first token) of domestic models kept rising this year, reaching 1.26 seconds in August, clearly up from the 2025 average of 0.78 seconds. But the average TTFT of typical foreign models in August reached 2.08 seconds, also clearly up from 1.04 seconds in April.

Notably, the rise in TTFT is not simply a performance regression: it comes together with larger model parameter counts, longer context, higher call volumes and other factors.

The report finds that after two years of price war, model service prices rose markedly in 2026 at home and abroad. The domestic normalised average price was 6.8 yuan in the second half of 2025 but 13.46 yuan this year — nearly double. Foreign prices remain higher than domestic ones: the average normalised price of typical models released since 2026 is about 117.8 yuan.

Among the monitored models, 47.8% have a context length of 1 million tokens, and 95.7% are at 200K or above — more than 76 percentage points higher than at the end of 2025 — while the share of models below 200K keeps shrinking.

In May the report added new versions of four open-source models to its monitoring and found that TPS kept rising while TTFT kept falling. In August, for example, DeepSeek-V4-Pro’s TPS had reached twice that of DeepSeek-V3.2.

On latency, after the main models were upgraded in May, TTFT generally rose for a period, then gradually fell back to pre-upgrade levels as inference optimisation measures were implemented.

At the same time, the four models perform differently across platforms in TPS, TTFT and call success rate. Some third-party platforms clearly outperform the model services offered by official platforms, and the TPS gap for the same model between platforms can approach a factor of two. Different inference acceleration, optimisation strategies and cost control at each platform are the main reasons for the divergence.