AI Agents Now Burn 5x More Tokens Than Humans on OpenRouter

AI agents are consuming dramatically more tokens than human users on OpenRouter, according to new data reported by Tom's Hardware. As of August, agent usage reached 7.3 trillion tokens, compared to 1.4 trillion for humans — a fivefold difference.

The surge is driven largely by cached prompts. Andreessen Horowitz (a16z), citing OpenRouter, reports that more than 85% of agent tokens come from cached prompts, meaning agents repeatedly re-read old context rather than generating new output.

OpenRouter classifies API keys as agentic, mixed, or human using a 7-signal weighted composite score. The mixed category grew 4.7x over the same period, while agent token usage has climbed 14x since the February crossover point — when agents first surpassed human usage on the platform. Human usage grew 2.8x over that span.

The pattern extends beyond OpenRouter. A call center consultancy's September usage on Claude Code showed that 96% of all input was re-reading old conversation. In McKinsey's 2026 State of AI survey, 40% of respondents from large organizations reported scaling AI agents, up from 27% a year earlier.

The memory implications are significant. Micron expects RAM and storage shortages to worsen in 2027 and 2028, with customers paying more than this year. Futurum Group's Daniel Newman and Peter Walker have pointed to the growing token demands of AI agents as a factor in hardware demand.