# 智能体消耗token达人类5倍，OpenRouter议价力或减弱

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-08-23 08:47
- AIHOT 分数：44
- AIHOT 链接：https://aihot.virxact.com/items/cmt539auy04ibro73fkciheoq
- 原文链接：https://x.com/rohanpaul_ai/status/2091326423596404871

## AI 摘要

智能体token消耗量约为人类的5倍，自2月以来使用量增长约14倍，人类已成AI的少数用户。当智能体成为OpenRouter上token的主要来源，其通过比价压价模型供应商的能力可能减弱。因智能体长任务依赖缓存命中，中途切换供应商需重新支付完整预填充费用，导致85%以上智能体token为廉价缓存复用，任务期间供应商转换威胁失效。

## 正文

Agents are consuming tokens at nearly 5x the human rate, while their usage has exploded ~14X since February.

Once agents became the majority of tokens on OpenRouter, the router's ability to play suppliers off each other may be much less.

OpenRouter sits between developers and the companies that run the models. It makes money by price shopping, sending each query to the cheapest provider at the time.

When the next question is a single question, this is fine; there is no need for any continuity between the questions.

But, the behavior of an agent doing a long task is different: it resends the same long block of background instructions at each stage (i.e. cache hit).

A cache hit only lives on the machine still holding the warm prefix, and rerouting mid-task will mean paying the full pre-fill over again.

i.e. the model provider stores that cache-hit block in memory and charges only a small fraction to reuse it.

Which is why more than 85% of agent tokens on OpenRouter are these cheap reuses (cached prompt) rather than fresh ones.

The copy is stored on the servers of a single company . If the job is transferred to a cheaper competitor halfway through , the stored copy is discarded and the full block is repaid .

That will mean the agent remains with the provider they started the task with until the task is finished and the threat from the router to take their business elsewhere is eliminated.

So looks like the discounts routers can squeeze out of model providers many shrink on agent traffic well before they shrink anywhere else.

### 引用推文

> a16z：Humans are the minority user of AI Agents burn nearly 5x the tokens people do, up 14x since February Charts of the Week: https://www.a16z.news/p/charts-of-the-w...