AI demand is no longer coming only from people typing prompts into chat windows. On OpenRouter, a growing share of token usage is being generated by AI agents that call models while completing tasks.
According to OpenRouter analyst Peter Walker, February 6, 2026, may have marked a turning point: the last day when humans consumed more tokens than AI agents on the platform. Since then, agentic token usage has grown 14x, compared with a 2.8x increase in human usage.
Why agentic token usage is rising
The core shift is simple: AI agents do not always stop after one model response. They can continue working across longer stretches, calling AI systems repeatedly as they plan, check, revise, and continue a task.
That behavior changes the shape of demand. A person may ask for one answer and then decide what to do next. An agent, by contrast, can trigger additional AI processes along the way, expanding the amount of model work needed before the final result is complete.
This is why agentic token usage can grow much faster than human usage. The user may still be a person at the start of the workflow, but the ongoing token consumption increasingly comes from AI systems coordinating with other AI systems.
The cost picture is more complicated
A 14x increase in agentic token usage sounds like a direct warning about exploding AI costs. The source article makes clear that the raw token figure does not tell the full story.
Nearly 70 percent of agent token usage comes from cached prompts. These prompts are billed at much lower rates, which means the actual cost increase is not moving as quickly as the total token count.
That distinction matters for anyone tracking AI infrastructure demand. Tokens show how much model activity is happening, but billing depends on what kind of tokens are being processed and whether prompts can be reused from cache.
For agent workflows, caching can soften the economic impact of repeated model calls. The agents may be using many more tokens, but a large share of those tokens falls into a lower-cost category.
OpenRouter’s model mix shapes the numbers
OpenRouter has its own platform profile, and that affects how its token data should be read. The source notes that OpenRouter skews toward open-weight models.
Those models tend to be less token-efficient than models from OpenAI or Anthropic. In practical terms, that means comparable tasks may require more tokens on the kinds of models that are heavily represented on OpenRouter.
Even with that caveat, the trend is not presented as isolated to one platform. The source says the pattern likely looks similar at the major labs. That makes the OpenRouter data useful as a signal for a broader movement in AI usage: more model activity is being initiated by agents rather than directly by humans.
Reasoning models helped start token inflation
The rise in agentic usage is part of a wider token inflation pattern. The source article points to reasoning models as an earlier driver.
Reasoning models use more tokens because they spend longer processing before giving an answer. That can increase token usage even when the extra thinking is not necessary for the task.
Agentic systems add another layer to this dynamic. Instead of one model thinking longer, an agent may create a chain of model activity over time. Each step can add to the total token count, especially when the agent keeps working independently.
What this means for AI demand
The main implication is that AI systems are becoming a larger customer for AI infrastructure. Human users still begin many workflows, but agents can become the main consumers of tokens once a task is underway.
This changes how AI growth should be understood. Usage is not only about more people using AI products. It is also about AI products using other AI processes more often and for longer stretches.
For platforms, labs, and customers, the key metrics are no longer just user prompts and final responses. Agentic token usage, cached prompts, token efficiency, and model selection all shape the real cost and scale of AI activity.
OpenRouter’s data points to a future in which AI demand compounds through automation. The more agents are asked to work independently, the more they may rely on additional model calls to complete the work. That is the practical meaning behind the 14x jump: AI is becoming a major driver of its own usage.