이미지: METAL LAB 생성
Summary
- On OpenRouter, AI agent token consumption has grown 14-fold since February 6, 2026, far outpacing the 2.8x growth seen in human token use.
- About 70% of agent tokens come from cheaper cached prompts, so actual cost growth has been much more modest than the jump in usage.
- OpenRouter skews toward open-weight models, which tend to be less token-efficient, but a similar shift could show up at major labs too as reasoning models spread.
- 분기점
- 2026년 2월 6일, 사람이 에이전트보다 토큰을 더 많이 쓴 마지막 날로 추정
- 에이전트 토큰 증가
- 14배 (0.51조 → 7.3조 토큰)
- 사람 토큰 증가
- 같은 기간 2.8배
- 캐시 프롬프트 비중
- 에이전트 토큰 소비의 약 70%
- 데이터 출처
- 오픈라우터 애널리스트 피터 워커
- 오픈라우터 모델 특성
- 오픈웨이트 모델 비중이 높아 토큰 효율이 상대적으로 낮음
A Shift in Who's Actually Using the Tokens on OpenRouter
OpenRouter, the platform that routes API requests across multiple AI models, just surfaced a striking data point. According to figures compiled by OpenRouter analyst Peter Walker, February 6, 2026 was likely the last day humans consumed more tokens than AI agents did on the platform. Every day since, agent token consumption has outpaced human consumption.
AI agents work differently from chatbots, which just answer a question and stop. Agents search on their own, generate files, move on to the next task, and along the way keep calling on other AI processes to help out — often working independently for extended stretches. Because the work gets broken into so many steps, token consumption scales at a pace that looks nothing like a typical human conversation.
14x vs. 2.8x: The Gap Behind the Numbers
Since February 6, 2026, the gap between agent and human token growth has widened to more than fivefold.
| Category | Token Consumption Growth (since Feb 6, 2026) |
|---|---|
| AI Agents | 14x (0.51 trillion → 7.3 trillion tokens) |
| Humans | 2.8x |
It's worth noting that human token consumption kept rising over the same period too — it didn't decline. But agent growth has been so much steeper that the vast majority of tokens flowing through OpenRouter now represent AI systems calling other AI systems, not humans making requests.
Cached Prompts Are Keeping Costs in Check
A 14x jump in agent token consumption doesn't translate into a 14x jump in cost. That's because roughly 70% of all agent tokens come from cached prompts. Cached prompts reuse stored results from previous computations instead of recalculating them from scratch, and they're billed at a much lower rate than freshly computed tokens. As a result, the report notes, actual cost growth has been far more moderate than the surge in usage would suggest.
Could the Same Pattern Show Up at OpenAI and Anthropic?
OpenRouter leans more heavily on open-weight models than some other platforms do. The report points out that open-weight models tend to use more tokens than OpenAI's or Anthropic's models to complete the same task. Even so, the broader trend of rising token consumption is likely to show up at major labs as well. In fact, this shift already started with the arrival of reasoning models — models that "think" before answering. That extra thinking step means they sometimes burn through more tokens than necessary, even on questions that don't really call for extended reasoning.
Editor's Take
In some ways, it makes sense that agents would consume more tokens than humans. A chatbot answers a question and stops there, but an agent has to plan, call tools, and check results over and over until a task is actually done — and every one of those steps adds up in tokens. What's notable is that this pattern is already showing up in benchmarks, too. It echoes what we recently saw when NVIDIA wrapped Claude Opus 5 in its own custom harness and pushed its ARC-AGI-3 score from 30% to 100%: the performance gain came less from the model itself than from how long and how many steps the surrounding harness let it work through. If using more tokens keeps translating into higher scores, agent token consumption is more likely to keep climbing than to level off.
From a practical standpoint, this is worth keeping in mind for companies budgeting for agent deployments. If you estimate API costs based purely on raw token counts, you risk overestimating the budget by missing out on caching discounts. On the other hand, tasks that require fresh context each time — where caching doesn't kick in — will cost exactly what the token count suggests, so it's worth checking your cache hit rate during any early pilot phase.
If there's one forecast worth making, it's that next quarter's OpenRouter report will likely show agent tokens outnumbering human tokens by well over tenfold. By then, "token consumer" may bring AI to mind before it brings humans to mind at all.
Comments