February 6, 2026, marked a quiet turning point in the AI ecosystem: it was the last day humans consumed more tokens than AI agents on OpenRouter. Since then, agentic usage has exploded 14-fold, reaching 7.3 trillion tokens by late July, while human consumption grew just 28%. The machine is literally becoming its biggest customer.
The Cost Illusion: The Role of Caching
Raw numbers suggest out-of-control inflation, but the economic reality is more nuanced. Nearly 70% of agentic tokens come from cached prompts, billed at significantly lower rates than fresh queries. This mechanism amortizes the financial impact, explaining why invoices aren't growing at the same pace as volumes. This dynamic aligns with previous analyses on agent costs, where routing efficiency and caching emerged as critical levers to contain expenses.
Model Efficiency and Market Trends
OpenRouter, a platform aggregating hundreds of models, has a strong skew toward open-weight systems. These tend to be less token-efficient than proprietary models from OpenAI or Anthropic, making the data even more indicative. The phenomenon is not isolated: the growth in agentic consumption reflects a structural shift, where agents operate for longer periods and autonomously spin up additional AI processes. This aligns with AMD's data center strategy, which predicts a hardware architecture overhaul to support persistent agentic workloads.
Strategic Implications
The shift from human interactions to agentic flows continues to erode the market share of US giants, as highlighted by the offensive of Chinese models on OpenRouter. In this scenario, model choice depends less on response quality and more on inference economics at scale. With the potential acquisition of OpenRouter by Stripe, control over these token flows becomes a central strategic asset for the global AI infrastructure.

AI-generated comment
AI-generated comment
AI-generated comment
AI-generated comment