CAICT Report: Domestic large model token prices doubled compared with the second half of last year, average daily consumption up 3x
The Institute of Artificial Intelligence of the China Academy of Information and Communications Technology (CAICT) and other institutions jointly released the "Large Model Token Service Performance Monitoring Report (2026)" on October 23, showing that domestic token prices doubled compared with the second half of last year, with the normalized average price rising from 6.8 yuan to 13.46 yuan, while average daily consumption grew 3x, reaching about 500 trillion in June 2026, nearly 4 times that of December 2025. The report is based on monitoring data from January to August 2026; in August, the average TPS of domestic models reached 78.23, double the 2025 average of 38.25, and the average time to first token rose to 1.26 seconds; typical foreign models had an average TPS of 112.82 and an average TTFT of 2.08 seconds in August. The report also compares four open-source models, DeepSeek, Qwen, GLM, and Kimi, noting that the TPS gap for the same model across different platforms can be close to a factor of two.
Why read this
AI score 59
The report uses normalized prices and monitoring data such as TPS and latency to present the simultaneous rise in volume and price of large model token services and the gap between domestic and foreign models.
Topics
Body · AI translation
Article body
With the rapid iteration of large model technology, Token services have become a key node in the transition of technology toward application. On October 23, the Institute of Artificial Intelligence of the China Academy of Information and Communications Technology, 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 jointly released the "Large Model Token Service Performance Monitoring Report (2026)" (hereinafter referred to as the report).
The report shows that token services present a trend of rising volume and price, among which domestic token prices doubled compared with the second half of last year, but volume growth reached 3 times. In June 2026, domestic average daily token consumption was about 500 trillion, nearly 4 times that of December 2025. More and more Model-as-a-Service (MaaS) platform vendors are supplying large model token services through public cloud APIs, and the lack of intuitive criteria for users to judge token service performance and the difficulty they face in model selection are the background for this report.
Based on monitoring data from January to August 2026, the report analyzes domestic and foreign large model token services across dimensions such as invocation success rate, throughput speed, first-token latency, stability, price, and context window.
After the price wars of the past two years, model service prices at home and abroad rose significantly in 2026: the normalized average price (note: converting different pricing methods of various vendors onto a single unified basis so that prices can be compared horizontally) rose from 6.8 yuan in the second half of 2025 to 13.46 yuan this year, an increase of nearly one fold; however, foreign prices remain significantly higher than domestic ones, with the average normalized price of typical models released since 2026 at about 117.8 yuan.
In terms of performance, in August the average output tokens per second (TPS) of domestic models reached 78.23, doubling the 2025 average of 38.25; but in the same period the average TPS of typical foreign models was 112.82, still overall better than domestic ones. In terms of latency, the average first-token latency (TTFT) of domestic models has shown a continuous upward trend, reaching an average of 1.26 seconds in August, a marked increase from the 2025 average of 0.78 seconds; the average TTFT of typical foreign models in August was 2.08 seconds, also showing an upward trend. The report explains that the rise in latency is not simply performance degradation, but the result of multiple factors occurring simultaneously, including increases in model parameter count, context length, and invocation volume.
Among the models monitored in the report, models with a context length of 1 million (1M) tokens already account for 47.8%, and those with 200,000 (200K) tokens and above account for 95.7% in total, an increase of more than 76 percentage points compared with the end of 2025, while the share of models with shorter context continues to decline.
The report also compares the performance of four open-source models—DeepSeek, Qwen, GLM, and Kimi—on different third-party platforms. After the four open-source models completed new version upgrades in May, TPS continued to improve and latency continued to compress—based on August data, the TPS of DeepSeek-V4-Pro has reached twice that of DeepSeek-V3.2. But the same model performs differently across platforms: the performance of some third-party platforms is significantly better than that of official platforms, and the TPS gap for the same model across different platforms can be nearly one fold at maximum. The report believes that differences among platforms in inference acceleration, optimization strategies, and cost control for the same model are the main reasons for the performance divergence.
(Source: The Paper) Middle of the body Embedded advertisement