OpenRouter data shows weekly token consumption rose over 25,000 percent since January 2025, moving from 0.5 trillion to 126.2 trillion units. This metric tracks the basic processing steps for AI models rather than direct business output. The sharp increase reflects inflated token counts rather than a proportional jump in actual user activity or economic value. Reasoning models generate extensive internal thought processes before delivering answers, which dramatically expands the total token tally. Unoptimised agentic systems often burn through these resources at high rates, causing small usage shifts to appear as massive spikes in consumption figures.
The chart highlights how specific model architectures distort standard performance measurements. OpenAI‘s GPT 5.6 Luna recently topped the consumption list, yet this dominance stems from its tendency to produce more tokens per prompt rather than wider adoption. Revenue leaders like OpenAI’s Astra show different trends, while Chinese models such as Kimi, GLM, and DeepSeek have increased monthly spending tenfold in 2026 from a smaller starting point. The data suggests current usage statistics may not accurately reflect real-world utility or sustainability.
- Reasoning models inflate counts via extensive internal processing steps.
- GPT 5.6 Luna leads volume due to high output per prompt.
- Chinese models show rapid growth from a lower baseline.
