CAICT: China's domestic model prices rose almost 100% in 2026 as daily token use hit 500 trillion

The China Academy of Information and Communications Technology (CAICT) says the normalized average price of domestic models reached 13.46 yuan so far in 2026, up from 6.8 yuan in the second half of 2025, an increase of almost 100 percent. The increase follows two years of price cuts, according to the report, as reported by Pandaily.

CAICT produced the 2026 large-model token service performance report with a Ministry of Industry and Information Technology key laboratory for large-model and hardware-software evaluation and with the Artificial Intelligence Industry Alliance. It uses monitoring data from January to August 2026. It compares official model platforms with third-party providers serving four open-source models: DeepSeek, Qwen, GLM and Kimi.

Key figures

• Usage: Daily token consumption in China reached about 500 trillion in June 2026, nearly four times the level in December 2025.
• Price: The normalized average price of domestic models was 6.8 yuan in the second half of 2025 and 13.46 yuan so far in 2026. Typical models from other markets launched in 2026 averaged about 117.8 yuan on the same normalized basis.
• Throughput: Average output throughput for domestic models was 78.23 tokens per second in August, about double the 2025 average of 38.25.
• Latency: Average time to first token was 1.26 seconds in August, up from 0.78 seconds in 2025.
• Context windows: Among monitored models, 47.8 percent support 1 million tokens and 95.7 percent support at least 200,000. CAICT says the latter share is more than 76 percentage points higher than at the end of 2025.

Why latency rose

CAICT says the higher latency is not simply a decline in performance. Larger models, longer contexts and higher call volumes all pushed latency up at the same time. The report found the same trend in typical models from other markets.

Platform differences

Throughput for the four open-source models kept rising after their May version updates. In August, DeepSeek-V4-Pro reached twice the throughput of DeepSeek-V3.2.

The same model performed very differently depending on where it ran. Throughput, latency and success rate varied widely between platforms, with throughput gaps of nearly 2x. Some third-party platforms outperformed the official ones. CAICT attributes the gaps to differences in inference acceleration, optimization and cost control.

More model-as-a-service platforms now sell tokens through public cloud APIs. The report says users still lack clear ways to judge service quality and often struggle to choose a provider.

All figures above come from the CAICT report.

What this means

The report shows that a model's name does not determine the service a buyer receives. CAICT found throughput gaps of nearly 2x for the same model across platforms, and some third-party platforms beat the official ones, so provider choice affects speed, latency and success rate directly.

The price data also marks a change in direction: domestic prices rose almost 100 percent after two years of cuts. CAICT frames the latency increase as a result of larger models, longer contexts and higher call volumes rather than weaker performance. Throughput roughly doubled in the same period. Teams buying tokens will need to track price, throughput, time to first token and success rate together rather than any single measure.