Throughput doubled, first token slower

The stamp at the top of this IT Home page is 2026/10/23 11:51:58. The AI institute of the China Academy of Information and Communications Technology, a ministry key laboratory for large-model and hardware-software evaluation, and the China Artificial Intelligence Industry Alliance jointly released the 2026 large-model token-service monitoring report on the day the page calls today, October 23. The report uses monitoring from January through August 2026. The page also says domestic daily token use in June 2026 was about 500 trillion, nearly four times the December 2025 level. Daily use is a volume figure. It does not stand in for August throughput, time to first token, or price.

Throughput is two sentences, not one doubling. The report says average domestic output in August reached 78.23 tokens per second, double the 2025 mean of 38.25. The same passage says the August mean for typical foreign models was 112.82 tokens per second, more than double their December 2025 figure of 51.35, and still ahead of the domestic mean. The 78.23 compares domestic models with their own 2025 mean. The 112.82 compares typical foreign models with their own December 2025 figure. The baseline months differ, and so do the levels. The domestic doubling does not put the domestic mean above the foreign one.

Time to first token is separate as well. The domestic average kept rising this year and reached 1.26 seconds in August, clearly up from the 2025 average of 0.78 seconds. Typical foreign models averaged 2.08 seconds in August, clearly up from 1.04 seconds in April. The foreign comparison month is April, not the 2025 average, so it is not the same pair as 0.78 against 1.26. The page says the rise is not simply a performance decline. It arrived together with more parameters, longer context, and more calls. That is the report's account of the rise, not a procedure someone could follow.

Price is separate too. The page's normalized domestic average was 6.8 yuan in the second half of 2025 and 13.46 yuan so far this year, nearly double. This story keeps the page's name for that average and does not invent the formula. Foreign prices remain higher. Typical models released since 2026 average about 117.8 yuan on the same normalized price. That is another price level, not the domestic move from 6.8 to 13.46. Among monitored models, 47.8 percent reach a 1 million token context, and 95.7 percent reach 200 thousand or more, up more than 76 percentage points from the end of 2025. In August, DeepSeek-V4-Pro throughput reached twice DeepSeek-V3.2. That is one lab's later model against its earlier one, not the domestic mean against the foreign mean. The same model can differ by nearly two times across platforms, and some third-party platforms beat the official service. That gap is between platforms. It does not fold into 78.23 or 112.82.