On October 23, the Artificial Intelligence Institute of the China Academy of Information and Communications Technology (CAICT) and its partners released the Large Model Token Service Performance Monitoring Report (2026). The number drawing the most attention is price: the normalized average price of China's LLM services rose from 6.8 yuan in the second half of 2025 to 13.46 yuan so far in 2026, an increase of nearly 100%.
The report was issued jointly by the CAICT AI Institute, the MIIT Key Laboratory for Evaluation of Large AI Models and Software/Hardware, and the China Artificial Intelligence Industry Alliance. Its data comes from monitoring between January and August 2026, covering call success rate, throughput, time to first token, stability, price and context window.
Price, speed and latency: three lines moving at once
Start with price. The report says that after two years of price wars, prices both in China and abroad rose noticeably in 2026. Typical models released overseas since the start of 2026 carry an average normalized price of about 117.8 yuan, still far above the domestic level.
Then speed. Domestic models averaged 78.23 tokens per second (TPS, output tokens per second) in August, against a 2025 average of 38.25, roughly a doubling. Typical overseas models averaged 112.82 TPS in August, up from 51.35 last December, also more than double, so the domestic field still trails.
Latency moved the other way. Domestic time to first token (TTFT) averaged 1.26 seconds in August versus 0.78 seconds across 2025. For typical overseas models, August TTFT was 2.08 seconds, up from just 1.04 seconds in April. The report's explanation:
"The rise in TTFT is not simply a performance regression."
It attributes the increase to a combination of larger parameter counts, longer contexts and heavier call volumes. The shift in context length is easy to see: among the monitored models, those with a 1-million-token context make up 47.8%, and models with 200K or more account for 95.7% combined, more than 76 percentage points higher than at the end of 2025.
A back-of-the-envelope tally
Putting the figures side by side yields a few rough ratios.
- Price: 13.46 ÷ 6.8 ≈ 1.98x; the overseas average is about 117.8 ÷ 13.46 ≈ 8.8 times the domestic one.
- Speed: domestic TPS rose about 2.05x, almost in lockstep with the price increase.
- Latency: domestic time to first token grew by about 0.48 seconds, an increase of roughly 60%.
If you estimate output speed per yuan spent, this year looks roughly flat against last year: the price doubled, and so did throughput. What developers are paying extra for is mainly longer context and larger models, while the wait for the first token on each request has actually gotten longer. This is our own estimate based on the averages the report discloses; individual models and time slots vary widely, so it should not be applied to any single vendor's bill.
Price moves by domestic vendors this year add some color to these averages. Tencent Cloud has raised prices for Hunyuan and other AI services several times this year. According to media reports, the DeepSeek-V4 API has used time-of-day pricing since August 17, with peak-hour output prices rising to 4.5 times the earlier level. Prices that were pushed to the floor during the price war are being walked back, one vendor at a time.
Same model, different platforms
The report also tracked four open-source models, DeepSeek, Qwen, GLM and Kimi, on their original vendors' platforms and on third-party platforms. It found that TPS for the same model can differ by close to a factor of two between platforms, and that some third-party platforms run faster than the official one. The report points to differing choices in inference acceleration, optimization strategy and cost control as the main source of this performance gap.
For individual models, DeepSeek-V4-Pro reached twice the TPS of DeepSeek-V3.2 in August. After the main model generation change in May, time to first token generally rose for a period, then gradually fell back to pre-upgrade levels once inference optimizations landed.
Growing usage explains why vendors dare to raise prices. The report notes that average daily token consumption in China reached about 500 trillion in June 2026, nearly four times the level of December 2025.
Readers should keep two caveats about the methodology in mind. The public summary does not explain how the "normalized price" is calculated or which pricing unit it uses, so the 6.8, 13.46 and 117.8 yuan figures can only be compared with each other on the same basis, and should not be set directly against any vendor's official list price. The summary also does not list the monitored models and platforms; the full report is the authority on that.
Sources: CAICT, CocoLoop, Jiemian News, ITHome, 21st Century Business Herald; the average normalized price, TPS, TTFT and context-share figures were checked against their stated statistical periods.