OpenRouter Eats 790 Trillion Tokens in October While AI Agents Take Over Everything

OpenRouter, the universal matchmaking service for various AI brains through a single API, decided to show everyone its utility bills for January through October 2026. Just in October, the platform processed an mind-boggling 790 trillion tokens, marking a twenty-five-fold explosion compared to January and a 37% jump from September alone.

Behind this massive digital appetite are AI agents, which now consume a whopping 76% of all traffic, translating to roughly 600 trillion tokens in October. Back in the ancient days of a year ago, these autonomous little helpers accounted for a meager 8% of the workload. The platform categorizes this traffic based on how API keys behave, looking at tool call frequencies, long interaction chains, and intervals between requests. Because input and output tokens are both tallied, this 76% figure doesn’t mean agents are inventing endless oceans of unique prose, as long sessions often involve feeding the exact same context back into the model repeatedly.

Caching mechanisms also got remarkably smarter, with the median hit rate among providers climbing from 76.1% in June to over 90% in October. Meanwhile, open-source models took a temporary gut punch when heavyweights like Opus 5.5, Sonnet 5.5, and GPT-6.1 Sol dropped on September 28, dragging the open-source market share down to 52%. However, the underdog models bounced back above 70% in just two weeks, largely thanks to the heroic intervention of DeepSeek V4.1 Flash.

Economics are finally starting to look somewhat merciful as the average cost per million tokens continues its downward spiral for both proprietary and open variants. Since January, actual spending on closed models plummeted by roughly 57%, while open models became 70% cheaper per million units. This overall cost reduction happens despite individual models occasionally hiking their prices, as caching efficiency and shifting usage patterns drag the platform average down. Leading the charge in sheer volume growth since May are platforms like Z.ai and DeepSeek, whose monthly token consumption multiplied by roughly 20 and 12 times respectively.

The sheer velocity of artificial intelligence scaling has officially transitioned from a quiet scientific experiment into an industrial-scale resource furnace, quietly reshaping global infrastructure while normal software developers just try to keep their databases from melting.

Source: OpenRouter Data