Tokenmaxxing: Is high AI usage a virtue or an expensive habit?

AI models count in tokens, small chunks of text that both what you type in and what the model replies are split into. The more tokens, the higher the bill. Now the term tokenmaxxing has begun to circulate, but it doesn't mean the same thing to everyone.

According to TechTarget, the word is used in two opposite ways. Either it means chasing high token consumption as a goal in itself or as a sign that you are using AI. Or it means getting the most possible value out of every token through better prompts, system design and choice of model.

Usage as a strategy

Blogger Tom Tunguz deliberately uses the term to describe maximizing his usage in order to become more productive. He writes that over six weeks he has increased his usage 20-fold and has used 250 million tokens in a single day. The figures are his own.

He points to research from METR showing that the latest models can work independently for 12 hours, compared with 1 hour a year earlier. His advice is to make a plan for the day in the morning so that several AI agents can work at the same time. One example is five agents on a single presentation: one analyzes git commits, one goes through error logs, one fact-checks sources, one builds the presentation, and one critiques it.

Usage as a trap

Thoughtworks sees it the opposite way and defines token-maxxing as the harmful practice of optimizing AI systems for higher usage rather than better results. The consultancy recommends governance and runtime control, fewer agents, and treating overuse as a question of how the system is built. Shayan Mohanty, Chief Data and AI Officer at Thoughtworks, says: "The challenge for enterprises is no longer just how to build agents, but how to govern the resources they consume."

i-SCOOP likewise argues that many tokens do not equal value. It can mean a higher price, slower systems, more noise and "context rot," where too much context degrades the model's focus.

Where the bill hides

TechTarget writes that the model's responses typically cost 2-4 times more per token than what you send in, and that the largest models are significantly more expensive than smaller ones. Four hidden cost drivers are long prompts containing entire documents, repeated injection of the same context, wordy responses with no length limit, and agent loops with no stop conditions.

Among the measures, TechTarget mentions sending tasks to an appropriately sized model, tightening context management, setting limits on retries and loop depth, and requiring structured output. An expert quoted by TechTarget estimates that model routing alone can cut token spending by 60% on mixed tasks. The figure is a single estimate with no stated basis.

i-SCOOP recommends starting with the desired outcomes rather than token budgets. TechTarget also points to security and compliance risks, including data leakage via prompts and missing audit logs.

TechTarget, Thoughtworks and i-SCOOP all produce advisory or consulting content with interests of their own.