蓝狐|8月 23, 2026 03:17
This a16z chart (data from OpenRouter) shows: in API/Agent routing traffic, humans are no longer the main driver of token consumption.
Agents currently consume about 5 times more tokens than humans, and since February 2026, their consumption has grown approximately 14-fold.
The reason lies in different work modes: multi-step planning, tool invocation, and repeatedly reading/writing context. Among these, cached tokens account for the majority of agentic tokens (a16z states it's over 85%). Humans tend to engage in Q&A interactions, while Agents lean toward continuous iteration.
This doesn’t mean humans have taken a backseat in consumer AI products, nor does it mean Agents have fully become foundational infrastructure.
What it indicates is: incremental demand is increasingly coming from autonomous Agents, not just from chat-based interactions.
Cost structures will be rewritten:
Token volume will rise, but caching will become cheaper;
The real winners of this incremental growth are likely those who make Agent loops efficient and can provide affordable, stable inference and tools.
For crypto, this represents a potential opportunity—possibly even a structural one.
For Agents to truly operate autonomously, they need to pay for compute, data, and tools themselves, rather than relying on humans to open API keys or swipe cards every time.
x402 offers an on-chain micropayment solution; low-fee L2s (like Base and other L2s) are better suited for such payments; Ethereum is exploring identity/reputation/verification standards (e.g., ERC-8004).
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink