On October 23, the Artificial Intelligence Research Institute of China's CAICT (China Academy of Information and Communications Technology), the MIIT Key Laboratory for Evaluating Large AI Models and Related Hardware and Software, and the AI Industry Alliance of China jointly released the Large Model Token Service Performance Monitoring Report (2026). Drawing on monitoring data from January to August 2026, the report lands on a clear verdict: after two years of price wars, domestic large-model token services have entered a phase of rising volume and rising prices. Daily token consumption in China averaged roughly 500 trillion in June 2026, nearly four times the December 2025 figure, while the normalized average price climbed from 6.8 yuan in the second half of 2025 to 13.46 yuan, an increase of close to 100%.

Why Tokens Need a Checkup Report

Tokens are the basic billing and measurement unit for large models. As MaaS (model-as-a-service) platforms deliver token services through public cloud APIs, the question users face is no longer whether a model is good, but which platform and which vendor's model get the same job done more cheaply, quickly and reliably. CAICT has run this monitoring effort since 2025. This round covers six dimensions — call success rate, throughput, time to first token, stability, price and context window — comparing first-party official platforms against the performance of four open-source models, DeepSeek, Qwen, GLM and Kimi, on third-party platforms.

Unit Prices Back to Pre-Price-War Levels, While Call Volume Quadruples

Price is the most counterintuitive finding. After two years of price wars, the report says, model service prices rose markedly in 2026 both in China and abroad. The domestic normalized average price went from 6.8 yuan in the second half of 2025 to 13.46 yuan so far this year, an increase of nearly 100%. For comparison, typical overseas models released since the start of 2026 carry an average normalized price of about 117.8 yuan. Normalization, per The Paper's explanatory note, converts each vendor's different pricing scheme into a common basis for side-by-side comparison. Among its core observations, the report describes the trend as "a clear price rebound, evolving toward a 'capability leap.'"

Volume matters more. Even with unit prices close to doubling, average daily token consumption in China reached roughly 500 trillion in June, nearly four times the December 2025 level — the market did not shrink as prices rose. The nature of demand has shifted: as models move into customer service, office work and R&D pipelines, users are paying for reliability, concurrency and context length, not just the price per token.

Throughput Doubles but Still Trails Overseas, and Rising Latency Isn't All Regression

A performance gap between China and overseas persists. In August, domestic models averaged 78.23 tokens per second (TPS), double the 2025 average of 38.25. Typical overseas models averaged 112.82 TPS over the same period, also more than double their 51.35 in December 2025, and better overall than the domestic figures.

Latency data looks like it is getting worse: domestic average time to first token (TTFT) stood at 1.26 seconds in August, up from 0.78 seconds in 2025. The report explicitly addresses this, noting that rising TTFT is not simply performance regression but a byproduct of simultaneous growth in model parameter counts, context length and call volume. Overseas models show the same pattern, with typical models averaging 2.08 seconds TTFT in August, a clear increase from 1.04 seconds in April. Context data backs this up: 47.8% of monitored models now support a 1M-token context window, and 95.7% support 200K or more, up more than 76 percentage points from the end of 2025. Requests are getting longer and heavier, so reading latency numbers in isolation is misleading.

Same Model, Different Platform, Almost Twice the Speed

Another set of findings concerns open-source models and third-party platforms. After the four open-source models went through a generational upgrade in May, TPS kept climbing and latency kept falling — as of August, DeepSeek-V4-Pro's TPS was twice that of DeepSeek-V3.2. But TPS, TTFT and call success rates for the four models were not consistent across platforms. Some third-party platforms clearly outperformed official ones, and for the same model the gap in TPS between platforms reached nearly a factor of two. The report attributes this to differences in inference acceleration, optimization strategy and cost control across platforms.

That pulls "token services" out of model benchmark scores and treats it as a product category to be evaluated in its own right: choosing a vendor means looking not just at model leaderboards, but at the service layer carrying the model.