Agentic Tokens Jumped 14x: What OpenRouter's Data Shows

February 6, 2026 may go down as the last day AI agents were the minority consumer. Since then, the machines have been paying the bills.

The Number

According to OpenRouter analyst Peter Walker, agentic token usage jumped from 0.51 trillion to 7.3 trillion tokens between February and August 2026 — a 14x increase over six months. Human token usage grew 2.8x over the same period. For the first time, AI is not just a tool that humans wield; it is increasingly a customer that other AI systems serve.

How the Math Works

The 14x figure compresses two compounding dynamics. First, agentic workflows are structurally token-heavy: agents spawn subagents, maintain long context windows, issue tool calls, and parse structured outputs — each loop multiplying token counts against the base workload. Second, the rise of reasoning models has extended per-task processing time even on tasks that previously resolved in a single pass, inflating token budgets without a proportional increase in completed work.

One moderating factor: approximately 70 percent of agentic token usage derives from cached prompts, which carry substantially lower billing rates than standard tokens. This means actual cost increases lag behind raw volume growth — but not indefinitely. As cache-ineligible operations scale up (dynamic tool outputs, multi-model routing, real-time retrieval), the discount compresses.

OpenRouter’s platform favors open-weight models, which typically consume more tokens than proprietary alternatives from OpenAI or Anthropic. Still, analysts expect similar consumption curves across all major providers — the trend is platform-agnostic.

What It Signals

The shift from human-driven to agent-driven consumption changes the unit economics of AI products. “Agents increasingly work on their own over longer stretches, spinning up additional AI processes,” Walker notes, meaning consumption is no longer bounded by human attention spans or working hours. Per-seat and per-query pricing models — built around human usage patterns — are increasingly misaligned with the actual cost structure of agentic deployments.

For platform and infrastructure teams, this signals that token budgeting needs to account for exponential growth trajectories, not linear ones. Stack Archive’s coverage of agentic CI systems hitting CI queue bottlenecks earlier this week illustrates the same compounding dynamic at the workflow layer: infrastructure designed for human throughput doesn’t hold at agent scale.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.