“Using Lots of Tokens Doesn’t Mean You Work Well”

In the rapidly evolving world of artificial intelligence, the term 'token' has emerged as a fundamental unit, akin to the 'coins' sought in the virtual world of 'Ready Player One.' As AI becomes indispensable in our daily lives – from learning and working to leisure – understanding tokens is crucial. China's recent rebranding of tokens to 'ciyuan' (词元), signifying 'word unit,' underscores its strategic intent to dominate the burgeoning token economy.

For those seeking a practical grasp of tokens, OpenAI's 'Tokenizer' offers a direct visualization. When we input text into generative AI, it is broken down into these discrete 'tokens.' Typically, four English characters constitute one token. The tokenizer reveals that a sentence like 'Tokens are becoming the digital rent' is parsed into 11 tokens (based on GPT-5). This illustrates the granular nature of AI processing and the associated costs.

“

Tokens are becoming the digital rent

An intriguing phenomenon observed in Silicon Valley earlier this year was the 'token maxing' culture, where token usage was perceived as a metric of developer productivity. However, research, such as the paper 'How AI Agents Use Your Money' by a joint team from the University of Michigan and Stanford, challenges this notion. Their findings indicate that token count can vary up to 30-fold for the same task, and crucially, the highest accuracy is achieved with moderate token usage. Over-reliance on tokens can lead to inefficient, repetitive actions, especially when AI models lack the mechanism to recognize and halt unresolvable tasks, as the paper points out.

This complexity makes pricing AI services a non-trivial issue. While technological advancements, like Google's 'Gemini 3 Flash' model, are improving token efficiency by reducing usage without compromising performance, the underlying economics remain a significant consideration. The discourse around tokens is not just technical; it touches upon economic models, efficiency, and the very future of how we interact with and pay for artificial intelligence.

“

AI agents do not have a mechanism to recognize that a task is unsolvable and stop early.