In this news:
The global artificial intelligence boom recently crossed a landmark threshold: autonomous agents now consume more than five times as much data as human users, triggering a surge in processing costs that is giving lower-priced Chinese models a competitive edge, according to analysts.
Unlike traditional chatbots that simply respond to human prompts, these agents operate independently to execute multi-step workflows, such as writing software, conducting research or managing business operations.
A single assignment can trigger a cascade of automated model calls as an agent plans its task, searches databases, invokes software tools and audits its own results.
This shift has dramatically altered the economics of AI deployment. Agentic requests consumed about 15 times more tokens – the basic units of data processed by an AI model – than standard human queries, according to data from model aggregator OpenRouter compiled by venture capital firm Andreessen Horowitz.
The data showed that daily token consumption from agentic workloads on OpenRouter hit 7.3 trillion in early August – 14 times the level six months earlier.
Agentic activity first surpassed human usage in February. By August, it accounted for over five times the 1.4 trillion tokens generated by human users on the platform.
Token usage will continue to grow exponentially as models become more capable and penetrate more scenarios
Sun Wenhao, TH Capital
Much of the increase came from cached prompt tokens – repeated portions of prompts, including system instructions, conversation histories and tool definitions, that can be reused without being fully processed again. Providers generally charge less for cached input because it requires less computation, although the rapid growth in overall volume could still increase corporate AI bills.
“Token usage will continue to grow exponentially as models become more capable and penetrate more scenarios,” said Sun Wenhao, a senior investment manager at Chinese deep tech venture investor TH Capital.
In China, daily token calls jumped 5,000-fold to 500 trillion in June this year from around 100 billion at the start of 2024, according to official data cited by state broadcaster CCTV.
To keep costs manageable, companies are increasingly turning to “model routing”, an approach where routine, repetitive tasks are directed to lower-cost models, reserving expensive, high-powered frontier models for complex reasoning, according to US enterprise software firm Salesforce.
“If everything runs on premium models, it is clearly too expensive,” Sun said, adding that coding and text-based research currently dominated agent usage, mostly assisting human workers rather than fully automating jobs.
Token demand could rise further once agents were trusted to carry out more tasks from beginning to end, he added.
The cost sensitivity could work to the advantage of Chinese open-weight models – AI systems that are released freely over the internet, allowing users with the necessary hardware to download and customise them for their own use.
For example, DeepSeek’s V4-Flash charges US$0.14 per million standard input tokens and US$0.28 per million output tokens. By comparison, OpenAI’s GPT-5.6 Sol charges US$4 and US$20, respectively, while Anthropic’s Claude Fable 5.1 costs US$10 and US$50.
DeepSeek accounted for 22.7 per cent of text requests on OpenRouter as of Thursday, the largest share of any model developer, ahead of Google’s 20.4 per cent and OpenAI’s 19.6 per cent, according to the platform’s live rankings.
However, cost is not the only hurdle facing widespread agent adoption.
Agents could generate code, analysis and documents faster than employees could vet them, potentially shifting the operational bottleneck from AI processing speed to human oversight, according to research from the MIT Sloan School of Management published in June.
.png)
