OpenRouter Tokens 7x in a Year. Agents Now Eat 5.6T a Week

part 1: the rise of the agent economy - what the data actually shows

LLM usage on @OpenRouter grew ~7x from ~3T to ~20T tokens a week in 12 months

but the volume isn't the story

a year ago, the top apps were chats and conversations

today: the top 4 are all agents - openclaw, kilo code, hermes agent, claude code

2/4 those 4 agents currently burn ~5.6T tokens a week between them

almost 2x what the entire openrouter platform consumed in a week in early 2025

leading the pack: @openclaw, the #1 app on openrouter at over 400B tokens a day. a year ago, it didn't exist

neither did @NousResearch's hermes agent, which went from near-zero to over 100B tokens a day in the last two weeks

3/4 what makes it all economically viable? chinese models

late 2024: chinese models were ~1.2% of all tokens on openrouter.
today: 45%+

minimax, kimi, deepseek, qwen - agent-native models built for tool-calling and long context, priced 15–20x cheaper than claude opus, with benchmark gap at almost 0

no cheap inference - no agent economy

4/4 the agent economy isn't coming. it's already running fast