China's AI Token Prices Nearly Double as Daily Use Reaches 500 Trillion

More AI use, higher prices and a longer wait to begin

China's daily consumption of artificial intelligence tokens reached about 500 trillion in June 2026, nearly four times the level recorded in December 2025, according to the China Academy of Information and Communications Technology (CAICT). Prices rose alongside demand: the average domestic price, adjusted to a common basis for comparison, increased from 6.8 yuan in the second half of 2025 to 13.46 yuan in 2026, a rise of about 98%.

CAICT's Large Model Token Service Performance Monitoring Report (2026), released on October 23, also records a change that users may notice directly. Domestic models generated text about twice as quickly in August as the average during 2025, yet took longer to produce their first token. Once an answer started, it arrived faster; before it started, the wait was longer.

The findings bring together a market recovering from two years of price competition, rapidly growing consumption and services able to process much larger amounts of information. They also show why choosing an AI service requires more than checking a model's name: different platforms running the same model delivered substantially different results.

What China is consuming at such scale

A token is a small unit of information that an AI model processes or generates. In text, it can be a word, part of a word or a punctuation mark. Tokens provide a common way to measure how much work a model handles, although their relationship to words varies across languages and systems.

CAICT's figure of about 500 trillion tokens a day therefore describes the volume of model activity. The near fourfold increase since December 2025 amounts to growth of roughly 300%, making the June total almost four times the earlier level.

More providers now sell access through model as a service platforms, often shortened to MaaS. Customers send requests through a public cloud application programming interface, or API, a connection that lets one piece of software ask another to perform a task. Businesses can use this route to put AI into their products without operating the underlying model themselves.

What the monitoring examined

CAICT's Artificial Intelligence Research Institute published the report jointly with the Ministry of Industry and Information Technology's key laboratory for large model and hardware and software evaluation, and the Artificial Intelligence Industry Alliance. Its monitoring covers January through August 2026.

The assessment examines request success rates, output speed, the wait before output begins, stability, prices and context windows. It compares official model platforms with third party services offering DeepSeek, Qwen, GLM and Kimi, four families of open source models.

CAICT says users often lack an easy way to judge token service performance and struggle to select a provider. Its measurements address that practical problem by separating several aspects of service that a single price or speed figure would leave hidden.

The main dates show how the changes unfolded:

• December 2025: the comparison point for the subsequent near fourfold increase in domestic daily token consumption.
• January to August 2026: the report's monitoring period.
• May 2026: monitoring incorporated new versions of the four open source model families.
• June 2026: domestic daily token consumption reached about 500 trillion.
• August 2026: domestic output speed averaged 78.23 tokens per second.
• October 23, 2026: the report was released.

Prices rebound after two years of cuts

Domestic model services entered 2026 after two years of price competition. The increase from 6.8 yuan to 13.46 yuan represents an additional 6.66 yuan on CAICT's normalized pricing measure, almost doubling the earlier average.

Normalization means converting different pricing arrangements to a common basis so they can be compared. The billing quantity behind these quoted yuan figures is unspecified, leaving customers without a direct conversion from the report's average to the cost of a particular request.

Typical models from outside China launched in 2026 averaged about 117.8 yuan on the same normalized measure. That is approximately 8.8 times the domestic average of 13.46 yuan. The comparison describes the monitored groups, while customers still face individual prices for each model and platform.

Answers arrive faster once generation starts

Domestic models produced an average of 78.23 tokens per second in August, compared with an average of 38.25 during 2025. The increase was about 105%, meaning output speed slightly more than doubled.

This measure, known as throughput or tokens per second (TPS), describes how quickly a service generates its answer. A higher figure can shorten the time needed to deliver a long response once generation is under way.

Typical models outside China averaged 112.82 tokens per second in August, about 44% above the domestic figure. Their output speed had also increased sharply, from 51.35 tokens per second in December 2025, a rise of approximately 120%.

The comparison periods are different: the domestic figure is measured against the average for all of 2025, while the foreign figure is compared with December. Both series show a large improvement in output speed, with the monitored foreign models retaining the higher August average.

Why the first response takes longer

Domestic models took an average of 1.26 seconds to produce their first token in August, up from 0.78 seconds during 2025. The increase of 0.48 seconds amounts to about 62%.

This waiting period is called time to first token, or TTFT. For someone using a chatbot, it is the pause between submitting a request and seeing the answer begin. Throughput measures the pace after that pause, so a service can improve on one measure while becoming slower on the other.

Typical foreign models took an average of 2.08 seconds to begin responding in August, compared with 1.04 seconds in April. Domestic models therefore began output 0.82 seconds sooner on average in August, even though their subsequent output was slower.

CAICT attributes rising initial delays to several changes happening together: larger models, longer contexts and more requests. Larger models require more computation; longer contexts give them more information to process; heavier traffic places more demand on the service.

Models can handle much more information at once

Among the monitored models, 47.8% supported context windows of 1 million tokens. A much larger share, 95.7%, could handle at least 200,000 tokens, an increase of more than 76 percentage points from the end of 2025.

A context window is the amount of information a model can work with during a request. It can include instructions, earlier conversation and documents submitted for analysis. A larger window gives users more room to provide information together, rather than dividing it among separate requests.

The figures show how quickly longer contexts spread across the monitored models. Nearly half offered capacity of 1 million tokens, while only 4.3% remained below 200,000. The increase of more than 76 percentage points means the share supporting at least 200,000 tokens had been below 19.7% at the end of 2025.

For a business processing lengthy documents, this expanded capacity changes what can fit into a single task. It also helps explain why the report examines capacity alongside waiting time: a model may be handling far more information before its first visible response appears.

The same model performs differently across platforms

DeepSeek, Qwen, GLM and Kimi services showed differences in throughput, initial response time and request success rates depending on the provider. For the same model, the gap in output speed between platforms approached a factor of two.

Some third party services performed better than the official platforms. CAICT attributes the variation to differences in inference acceleration, optimization strategies and cost control. Inference is the computing process through which a trained model produces an answer; providers can organize and optimize that work differently.

Version changes mattered too. In August, DeepSeek-V4-Pro achieved twice the throughput of DeepSeek-V3.2. After the May upgrades, the monitored models generally experienced an initial rise in waiting time, followed by a decline as providers introduced optimizations. Delays gradually returned toward their levels before the upgrades.

Customers choosing a service therefore have two related decisions: which model to use and which platform should run it. The report's findings make performance at the provider level a concrete part of that choice.

What the changes mean for customers

A business buying AI access now faces higher average domestic prices alongside faster output and much larger context windows. The most useful comparison depends on the work: an interactive chat places a premium on a quick start, while generating a lengthy answer makes sustained output speed more consequential.

A document analysis service may need a large context window, while an application handling repeated customer requests also depends on successful calls and stable performance. CAICT includes all of these measures because each affects a different part of the user's experience.

The report's combined picture is a market expanding in volume and capability while becoming more expensive on its normalized price measure. Domestic services retained a substantial price advantage over the monitored foreign models and began responding sooner on average in August. Foreign models generated tokens faster after output started. Across both groups, the growing wait for the first token shows the pressure accompanying larger models and heavier use.

Key Points

• China consumed about 500 trillion AI tokens a day in June 2026, nearly four times the December 2025 level.
• Domestic normalized prices rose about 98%, from 6.8 yuan to 13.46 yuan.
• Domestic output speed averaged 78.23 tokens per second in August, more than double the 2025 average.
• The domestic wait for a first token increased from 0.78 seconds to 1.26 seconds.
• Some 95.7% of monitored models supported context windows of at least 200,000 tokens.
• The same model's output speed varied by nearly a factor of two across platforms, with some third party providers outperforming official services.