About 500 trillion, and that multiple is usage

The stamp at the top of this IT Home page is 2026/10/24 22:41:49. It relays the China Academy of Information and Communications Technology: in June 2026, average daily domestic token use was about 500 trillion, nearly four times December 2025. That multiple is usage. It is not the later throughput line, and it is not price. The monitoring report is named as the 2026 large-model token-service performance report. The data window is January through August 2026. The comparison covers official platforms and four open models, DeepSeek, Qwen, GLM, and Kimi, as served on third-party platforms.

Throughput and time to first token stay on separate lines. The report says that in August the domestic models averaged 78.23 output tokens a second, against a 2025 mean of 38.25, a doubling on the domestic line. Typical foreign models averaged 112.82 in August, against 51.35 in December 2025, also more than a doubling, and the page says the foreign line is ahead overall. Time to first token is another line: 1.26 seconds for the domestic August mean, against 0.78 in 2025. Typical foreign models averaged 2.08 seconds in August, against 1.04 in April. The page says the slower first token is not only a loss of speed. Parameter count, context length, and call volume moved with it.

Prices are normalized averages, not one vendor's sticker. Domestic normalized price was 6.8 yuan in the second half of 2025 and 13.46 yuan so far this year, close to a doubling. Typical foreign models released since 2026 average about 117.8 yuan on the same normalized price, and the page says foreign prices remain higher. Both figures are the page's normalized prices. Context length is a fourth line: 47.8 percent of monitored models reach 1 million tokens, and 95.7 percent reach 200,000 or more, up more than 76 percentage points from the share at the end of 2025. Those 76 are percentage points.

The open-model example does not cover every model. The page says that in August, DeepSeek-V4-Pro's tokens per second had reached twice DeepSeek-V3.2. That is an example. The same model on different platforms can differ by nearly a factor of two in tokens per second. The page gives the reason as different acceleration, optimization, and cost control. That factor of two is between platforms, not the V4-Pro to V3.2 example. A vehicle-count item at the bottom is a different document and is not folded in.