CAICT Report: In 2026, Average Prices of Domestic Large Models Doubled, and Per-Second Throughput Also Doubled

Report says domestic large model average price doubled in 2026, with per-second throughput also doubling

2026-10-23 11:51 IT Home - Wang Miao

Thanks to IT Home readers Buyo Ya and Cat Sitting on the Piano for the tip-off!

IT Home, October 23: The Institute of Artificial Intelligence of the China Academy of Information and Communications Technology (CAICT), the Key Laboratory of Artificial Intelligence Large Models and Software/Hardware Evaluation of the Ministry of Industry and Information Technology, and the China Artificial Intelligence Industry Development Alliance today (October 23) jointly released the "Large Model Token Service Performance Monitoring Report (2026)."

The report states that with the rapid iteration of large model technology, Token services have become a key point in the transition from technology to application, and China's daily average Token usage has continued to climb. In June 2026, domestic daily average Token consumption was about 500 trillion, nearly 4 times that of December 2025.

An increasing number of MaaS (Model as a Service) platform vendors in the industry are supplying large model Token services through public cloud APIs. However, users lack intuitive criteria for judging Token service performance and face various problems such as difficulty in selecting services.

Based on monitoring data from January to August 2026, the report analyzes dimensions including call success rate, throughput speed, time to first token, stability, price, and context window for Token services across platforms. It compares the performance differences between the official platforms of original large model developers and four open-source models—DeepSeek, Qwen, GLM, and Kimi—on third-party platforms, further analyzes the key factors affecting Token service performance, summarizes core industry observations, and clarifies the next work plan, aiming to provide reference for users in selecting services and direction for the industry to optimize service capabilities.

1. Domestic Token service per-second throughput doubled compared with 2025, but remains lower than overseas

The report states that in August, the average TPS (tokens output per second) of domestic models reached 78.23 (tokens/second), doubling from the 2025 average of 38.25 (tokens/second). However, in August the average TPS of typical overseas models was 112.82 (tokens/second), an increase of more than double compared with 51.35 (tokens/second) in December 2025, overall outperforming domestic models.

Monthly average TPS of all large model Token services

2. Token service latency has increased, with consistent trends domestically and overseas

The report states that since the beginning of this year, the average TTFT (time to first token) of domestic models has shown a continuous upward trend, reaching an average of 1.26 seconds in August, a marked increase compared with the 2025 average of 0.78 seconds. At the same time, however, the average TTFT of typical overseas models reached 2.08 seconds in August, also a marked increase compared with 1.04 seconds in April.

It is worth noting that the rise in TTFT is not simply a performance degradation, but rather the result of multiple factors occurring simultaneously, including increases in model parameter count, context length, and call volume.

Monthly average TTFT of all large model Token services

3. Token service prices rebounded significantly in 2026

The report states that after the price wars of the past two years, model service prices rose significantly at home and abroad in 2026. The normalized average price in China was 6.8 yuan in the second half of 2025, while the average price so far this year is 13.46 yuan, an increase of nearly 1 times. However, overseas prices remain higher than domestic ones, with the average normalized price of typical models released since 2026 at about 117.8 yuan.

Trend chart of normalized prices at release of domestic original large models

4. Model context lengths continue to expand toward the million-token level

The report states that among the monitored models, those with a context length of 1M (1 million tokens) accounted for 47.8%, and those at 200K and above totaled 95.7%, an increase of more than 76 percentage points compared with the end of 2025, while the share of models below 200K continued to decline.

Share of domestic large model context lengths

5. Open-source model Token service performance improved significantly, with large performance fluctuations for the same model across different platforms

The report states that in May, new versions of four open-source models were updated for monitoring, and the results showed continuous TPS improvement and steadily reduced TTFT latency. For example, in August the TPS of DeepSeek-V4-Pro had reached twice that of DeepSeek-V3.2.

In terms of latency, after the main models completed generational upgrades in May, TTFT generally showed a phased increase, and as inference optimization measures were implemented, latency metrics gradually fell back to pre-upgrade levels.

At the same time, the four models showed differing TPS, TTFT, and call success rates across different platforms. Some third-party platforms performed significantly better than the model services provided by official platforms, with the TPS gap for the same model across platforms reaching nearly one fold at its maximum. Differences in inference acceleration, optimization strategies, and cost control for the same model across platforms are the main reasons for the performance divergence.