AI agents have become the largest consumers of model tokens on OpenRouter, according to figures shared by analyst Peter Walker. Agent-driven consumption rose from 0.51 trillion tokens in February 2026 to 7.3 trillion, an increase of roughly fourteen times.
Human-led usage grew about 2.8 times over the same period. Walker’s analysis suggests February 6 may have been the last day when people directly consumed more tokens than agents on the platform. The shift reflects workflows in which software keeps models working for longer periods and may launch additional AI processes to complete subtasks.
Raw token volume does not translate directly into an equally large bill. Nearly 70% of agent token consumption reportedly comes from cached prompts. Providers charge much less for cached input because previously processed context can be reused instead of computed again. OpenRouter also carries many open-weight models, which may use more tokens than tightly optimized proprietary systems, so its exact ratio should not be assumed to represent the whole market.
Even with those caveats, the direction matters for infrastructure planning. Model providers increasingly need to optimize for machine-generated demand: long contexts, repeated instructions, tool loops, and bursts of parallel requests. The typical customer is no longer always a person waiting for a single answer.