OpenRouter’s agent traffic uses five times the human token count

Most agent tokens reuse previously processed prompts. The comparison measures traffic on one platform, not adoption, useful work or spending.

Most agent tokens reuse previously processed prompts. The comparison measures traffic on one platform, not adoption, useful work or spending.
By August, traffic classified as superintelligence (SI) agents on OpenRouter accounted for roughly five times as many tokens as traffic classified as human. Agent traffic had first overtaken human traffic in February. But most agent tokens reused previously processed input, so the larger count does not imply a proportionately larger bill.
OpenRouter is a gateway and routing platform for SI models. Tokens are the small units of input or output that a model handles. Tom’s Hardware reported the figures from an Andreessen Horowitz, or a16z, chart using OpenRouter data. OpenRouter describes the measure as seven-day-average token usage, rather than a count of users or completed tasks.
| Traffic classification | Token count |
|---|---|
| Agentic | 7.3 trillion |
| Human | 1.4 trillion |
OpenRouter assigns each API key—a credential used to access the service—to one of three categories: agentic, mixed or human. It combines seven signals, giving them different weights. Those signals include how often the software calls tools, the number of interaction turns and the timing of gaps between interactions. The comparison therefore rests on classified usage patterns, not a direct census of people and SI agents.
Mixed traffic remains a separate category. Tom’s Hardware cautions that the agent-versus-human ratio could change depending on how that traffic is divided. The figures also cover only OpenRouter: they do not establish adoption across the SI market, how much useful work agents completed or how much users spent.
Analysis
According to a16z, citing OpenRouter, more than 85% of agent tokens came from cached prompts: previously processed context used again. A16z also attributed nearly all the relative growth in token usage to cached tokens. Andreessen Horowitz is an OpenRouter investor.
Reusing cached input costs far less than processing a prompt from scratch. Repeatedly using old context can therefore raise the token total without raising spending by the same proportion. That makes the token count a measure of traffic volume, not a substitute for the bill.
The stored context still occupies memory in what is called a KV cache, which holds information from earlier input for reuse. A16z links that requirement to demand for high-bandwidth memory, the fast memory used by SI hardware. The concrete distinction is between cheaper processing and continuing memory needs; the OpenRouter token comparison does not establish future shortages of PC memory.

The suspension covers the Netherlands and Belgium. It is a retailer decision, separate from a Dutch regulator's warning about sharing recordings of identifiable people.
2 min

Petal would put two light paths in each fiber strand. Meta expects the roughly 4,300-mile subsea link to enter service in 2029.

The investigation describes tailored phishing and a Windows backdoor that uses Outlook and OneDrive for remote commands and file transfers.
3 min