Agents overtook humans in token consumption on OpenRouter, and the gap keeps widening

On February 6, 2026, API keys classified as agentic consumed more tokens than human keys on OpenRouter for the first time. By the last week of September the ratio had reached seven to one in the agents' favor, with no sign of the trend slowing. The monthly classification relies on signals such as tool-call frequency, conversation turn count, and the speed of successive requests.

What an agent workload looks like

An agent does not ask a question and wait for an answer. It re-sends the entire context at every step: system prompt, tool definitions, files read, and the full history of tool calls up to that moment. Each step adds a small amount of new information and requests a short reply. The result: in September roughly 90% of agent tokens were cached prompt tokens, up from about 60% in January. Novel tokens made up most of the remainder, while output together with reasoning accounted for a sliver of about 1%. The picture for humans in chat is the reverse: novel tokens are still the majority, and output with reasoning reaches about 12%, because people want to read answers while agents want to take actions.

Cache changes the economics

Cached tokens are dramatically cheaper. In current frontier models from Anthropic, OpenAI, and Google, a cache read costs one-tenth of a fresh input token and one-fift