CAICT Report: Domestic Daily Average Token Consumption in June 2026 Was About 500 Trillion

China consumes 500 trillion tokens a day. Source: Semiconductor Horizons. Published: 2026-10-26 11:03. AI large-model-generated posters are nearly 4 times as many as in December 2025.

No matter how popular large models become, when it comes to the application side, what ultimately matters is how fast they compute, how fast they respond, and how much they cost. These three things are now all compressed into a single dimension: the token. A token is the smallest unit of measurement for information processed by a large model, and it is extending from a technical concept inside the model to a key interface for service invocation, cost accounting, and even commercial billing. The latest "Large Model Token Service Performance Monitoring Report (2026)" released by the Artificial Intelligence Research Institute of the China Academy of Information and Communications Technology (CAICT) presents a set of staggering figures: in June 2026, domestic average daily token consumption was about 500 trillion, nearly 4 times that of December 2025. Large models are moving from whether the technology can run at all toward how applications can afford to use it and use it well, and token services have become the key point in the transition from technology to application.

Supporting this torrent are a growing number of MaaS (Model-as-a-Service) platform vendors. Through public cloud APIs, they bring large model token services to users across thousands of industries. Open-source models have become almost standard on every platform. According to CAICT statistics, the DeepSeek, Qwen, and GLM series all have a 100% deployment rate on 22 mainstream domestic and international MaaS platforms, while Kimi and MiniMax also reach 95% and 86%, respectively. Model choices are more abundant than ever, and the barrier to invocation keeps falling.

What is driving this surge is not just conversational applications. New forms of applications represented by agents are pushing token consumption toward exponential growth, with a single task potentially consuming tens or even hundreds of times as many tokens as an ordinary chat. The deeper the application, the more sensitive users become to the performance, stability, and cost of token services, and model selection has upgraded from "as long as it works" to "fast, stable, and economical."

But an awkward problem follows. Token services are everywhere, yet users find it hard to judge which provider is more reliable. How fast is throughput, how long must one wait for the first token, how stable is the service, how expensive is the bill—these metrics that directly determine experience and cost have long lacked a unified yardstick. Choosing based on guesswork, gut feeling, and marketing has become an unavoidable hurdle for many developers.

To give users a visible ruler, the Artificial Intelligence Research Institute of CAICT, relying on the Key Laboratory of the Ministry of Industry and Information Technology for AI Large Models and Software/Hardware Evaluation, together with the Model Service Working Group and the High-Quality Token Service Working Group of the China Artificial Intelligence Industry Alliance (AIIA), has since 2025 continuously carried out performance monitoring of public cloud large model token services. This report is based on monitoring data from January to August 2026 and systematically surveys the token services of mainstream domestic and international platforms across dimensions such as invocation success rate, throughput speed, first-token latency, stability, price, and context window.

The monitoring results outline a clear industry trend. At the throughput level, in August the average output tokens per second (TPS) of domestic models reached 78.23, doubling the 2025 average of 38.25, but the average TPS of typical foreign models in the same period has reached 112.82, still overall better than domestic ones. At the price level, after the price wars of the past two years, model service prices at home and abroad clearly rebounded in 2026. The domestic normalized average price rose from 6.8 yuan in the second half of 2025 to 13.46 yuan so far this year, an increase of nearly 1 times, while the average normalized price of newly released typical foreign models is about 117.8 yuan, still an order of magnitude higher than domestic ones. Context capabilities also continue to expand: among monitored models, those with context lengths reaching the million-token level account for 47.8%, and those at 200K and above total 95.7%, an increase of more than 76 percentage points compared with the end of 2025.

Behind the price rebound is a shift in industry logic from competing on price to competing on capability. Evaluators point out that token service performance is improving overall, and leading platforms continue to optimize key indicators such as per-second throughput, already meeting the requirements of most general scenarios. At the same time, the overall latency of original-vendor large models has increased somewhat, which is not unrelated to the growth in model parameter counts. In terms of the service stability index, domestic token service stability remained at a good level above 80 points from January to April, and has fluctuated since May, mainly because after new versions of some popular models were released, throughput and latency varied sharply across different time periods.

For users currently selecting models, the value of this report lies in comparing the cards of each platform under the same ruler for the first time. Fast computing, stable responses, and worthwhile spending are no longer vague marketing talk but real metrics that can be quantified and reproduced. The report also reminds that token service performance is comprehensively affected by factors such as model parameter count, platform scheduling capability, and context window, and users should make trade-offs based on their own scenarios: high-frequency invocation scenarios value throughput and stability more, while interactive applications value first-token latency more. As agents and multimodal applications push token consumption to higher levels, the performance battle over token services will become increasingly fierce, and this ruler may well be becoming the new starting point for the industry's next round of competition.

This article is reposted from media reports or online platforms and represents the author's personal stance or views. Our reposting is solely for sharing and does not mean we endorse or agree with it. If the source attribution is incorrect or your lawful rights and interests are infringed, please contact customer service promptly, and as a neutral platform service provider we will promptly correct, delete, or handle it in accordance with the law. Comments: No user comments yet. Latest News More Related Recommendations

Samsung self-tests Exynos 2700 performance about 20% higher than Qualcomm's flagship chip 2026-10-26 11:48

Large quantities of GPUs arrive but cannot be powered on! Amazon exposes the power bottleneck in AI computing expansion 2026-10-26 11:15

AMD Ryzen 5 5500F, can it shake up the sub-100-yuan CPU market? 2026-10-26 11:13