Artificial intelligence is becoming one of its own largest customers. The South China Morning Post reported on September 4 that autonomous agents were consuming more than five times as many tokens as human users on the OpenRouter platform by August, altering the economics of running AI services and potentially favoring lower-priced Chinese models.
The change is driven by the way agents work. A conventional chatbot often produces one answer from one or several user prompts. An agent assigned a longer task can plan, call a model many times, search databases, invoke a tool, inspect a result, and repeat the process. Every step adds tokens, the units of text and data processed by a model. That makes the price of inference more important as agents move from demonstrations into recurring business workflows.
SCMP cited OpenRouter data compiled by Andreessen Horowitz showing that agentic requests used about 15 times as many tokens as standard human queries. Daily agentic token usage reached 7.3 trillion in early August, about 14 times its level six months earlier, according to the report. The data applies to OpenRouter, not all global AI usage, but it offers a clear sign that automated model calls are becoming a major source of demand.
Agent Workflows Multiply the Cost of Each Task
The largest difference between a chatbot and an agent is not simply intelligence. It is the number of steps a system takes after receiving an instruction. A human who asks for a summary may trigger a limited exchange. An agent told to research a market, write software, or manage a business process can make many model calls as it divides the work, chooses tools, validates outputs, and tries again after an error.
SCMP describes those workflows as combining planning, searches, software tools, and self-auditing. A single assignment can therefore create a chain of model interactions that is far longer than the final answer suggests. The report says agentic activity first surpassed human token use on OpenRouter in February. By August, human users accounted for 1.4 trillion tokens on the platform while agents consumed more than five times that level.
The jump has strategic implications for model developers. A provider that can offer a reliable answer at low cost may do well with individual users. A provider that can deliver reasonable performance across hundreds or thousands of automated calls may have an even larger opportunity if businesses build agents into internal systems. Price, throughput, context length, tool reliability, and error rates all influence the final bill.
This is why headline API prices can mislead. A model priced cheaply per million tokens can become costly if it repeatedly fails and needs a human to intervene. A more expensive model may reduce total spending if it completes a task in fewer steps. SCMP’s main point is that as agent workloads expand, the price of each additional call becomes harder for customers to ignore.
China’s model companies have been preparing for this environment through price competition and smaller, faster systems. In August, EastFrontier reported that Alibaba priced Qwen3.8-Max at US$2 per million tokens. That earlier price comparison illustrates why a rise in automated usage can matter more than a rise in ordinary consumer chatting.
Chinese Providers Can Compete on More Than a Sticker Price
SCMP said analysts believe lower-priced Chinese models could gain an advantage as agentic costs rise. That is a possibility, not a forecast. Chinese providers still have to show that their models work reliably in multilingual settings, comply with customer data requirements, integrate with common tools, and provide sufficient support for enterprise deployments.
Cost nevertheless matters because agents can be used at much higher volume than a human-facing assistant. A coding agent might inspect a repository, open files, run tests, and revise code. A research agent might query multiple sources and check the consistency of its findings. In each case, the model may need a long context window and repeated tool calls. A small difference in per-token pricing can become significant when multiplied across many tasks and users.
The Chinese market is already producing tools that focus on whether agents can finish real work. EastFrontier covered Baidu’s DumateBench evaluation for Chinese AI agents, which shifts attention from a model’s isolated answer to task completion in practical workflows. As agents take on multi-step tasks, customers will care about the combined cost of reasoning, tool use, verification, and recovery from mistakes.
Model design can affect those costs as much as pricing does. Smaller models may handle document classification or data extraction, while more capable ones may be needed for difficult planning or code review. Some systems can route parts of a workflow to different models, reserving expensive inference for steps that require it. The sources reviewed for this article do not establish which Chinese providers have the best routing systems, so that question remains open.
The advantage may also be geographical. A company choosing a model for an agent must consider where it is hosted, which language it supports, how it handles local enterprise data, and whether it can connect to the software the customer already uses. Low prices can attract trials, but sustained adoption requires an ecosystem of tools, documentation, and integration partners.
OpenRouter Data Is a Signal, Not a Global Census
The OpenRouter figures are powerful because they show the direction of agent demand, but they should not be overgeneralized. OpenRouter is a model-aggregation platform with a particular user base. Its seven-day average of 7.3 trillion agentic tokens in early August does not represent every consumer chatbot, enterprise API, internal company system, or government service worldwide.
The data also captures an early stage of adoption. Agents that use many tokens may be conducting complex work, but they may also be inefficient, poorly designed, or still in testing. Token use alone does not measure business value. A company needs to know whether an agent produces correct work, reduces human time, and delivers enough value to justify the infrastructure expense.
China’s provider ecosystem is building toward this question. EastFrontier reported that Tencent released Hy4 for China’s AI-agent race, while Alibaba and other companies are pairing models with office tools, cloud services, and industry applications. The business opportunity is not merely selling tokens. It is helping customers use those tokens to complete work at an acceptable cost and risk level.
SCMP’s report identifies a shift that could change how model competition is measured. The early race focused heavily on model capability, parameter counts, and benchmark scores. An agent-heavy market will also reward cost control, tool-use reliability, context management, and the ability to operate over long task sequences.
For Chinese models, low inference prices could become a material advantage if they are paired with strong agent systems and enterprise distribution. The evidence does not yet show a decisive winner. It does show why the economics of AI are moving beyond the price of a single answer. In a world where machines generate most token demand, the model that is affordable enough to call repeatedly may gain an important edge.