📌 DID YOU KNOW?
Researchers found that agentic AI coding tasks can consume up to 1000× more tokens than normal AI chats.
Not 10×.
Not 100×.
1000×.
What's surprising is that more tokens didn't consistently improve results.
In many cases, performance peaked and then plateaued—even as token usage kept rising.
The same task could also vary by up to 30× in cost between runs.
AI agents don't just generate answers. They read context, call tools, retry tasks, and execute workflows, accumulating massive token usage behind the scenes.
The biggest AI bill often isn't the output—it's everything the agent did before reaching it.
The emerging enterprise challenge isn't just model performance.
It's token governance.
Because more AI spending ≠ more intelligence.
Sometimes it's just more tokens.
Source: Stanford Digital Economy Lab / arXiv Research