OpenRouter’s latest usage data presents a visual shock: weekly token consumption has rocketed from 0.5 trillion in January 2025 to 126.2 trillion today. That is a 25,000 percent increase. On the surface, this looks like explosive adoption. Dig deeper, and it reveals something else entirely. The metric is broken. Tokens are no longer a reliable proxy for user engagement or business value because reasoning models have changed how AI processes information.
Reasoning models do not just answer. They think aloud. Before delivering a final response, these systems generate massive volumes of intermediate “thinking” tokens. A single prompt can trigger thousands of invisible computational steps. This inflates the count without necessarily improving the user experience. Unoptimized agentic AI systems make it worse. They loop, retry, and burn through tokens at staggering rates. A small uptick in active users can now look like a massive spike in consumption. The chart measures inefficiency as much as demand.
Model-specific trends highlight this distortion. OpenAI’s GPT 5.6 Luna currently dominates token volume on the platform. This does not mean it has the most users. It likely means it generates more tokens per prompt than its competitors. Revenue tells a different story. OpenAI’s Astra leads in actual spending, suggesting that high-token models are not always the highest earners. Meanwhile, Chinese contenders like Kimi, GLM, and DeepSeek are gaining ground. Their monthly spending has grown tenfold in 2026. The base remains small, but the trajectory is steep.
This divergence between volume and value forces a reckoning for developers and investors. Relying on raw token counts to gauge market health is now dangerous. It rewards verbosity over utility. Companies optimizing for cost will need to strip out unnecessary reasoning steps. Those chasing vanity metrics will keep burning cash. The real story here is not growth. It is waste. Until the industry standardizes how we measure useful output, these charts will remain misleading. We are paying for thought, not just answers. The question is whether that thought is worth the price.
