Tokenmaxxing is Goodharting

Coding agents and reasoning models let individuals consume many more LLM tokens than they could a year ago. It’s now easy for a single engineer to spend thousands of dollars in daily token usage. This is being actively encouraged through the recent memetic spread of “Tokenmaxxing” – the idea that if you consume more tokens, you’re more “AI native” and therefore producing more valuable output.

Tokenmaxxing is not The Way. Plainly, it’s a textbook instance of Goodharting. Token leaderboards come from an understandable short-term instinct to shift habits towards more AI usage, but direct optimization in this fashion inevitably overshoots into wasteful spending. Token-usage-as-target means token consumption ceases to be a useful metric.

Per-engineer token usage is, admittedly, useful as a diagnostic when engineers are dramatically and systematically underusing AI. However, the leaderboard version of token usage is likely actively harmful. This is analogous to how “lines of code merged” or “PRs merged” or “design docs written” can be interesting aggregate diagnostic numbers when used directionally, but obviously a leaderboard of “design docs written per quarter” would produce the wrong incentives.

Operationalizing this into predictions over the timespan of the next 6 months:

• Weak claim: Prominent executives / thought-leaders start publicly criticizing Tokenmaxxing / raw token usage as a bad metric.
  • This is basically already happening. My prediction: 90%.
• Stronger claim: There is a noticeable cross-company narrative shift toward discipline on AI spend with an emphasis on ROI measurement and skepticism of using raw token as a proxy for productivity.
  • I’m more tentative on this¹. My prediction: 60-70%.

Reasoning:

• There was a sharp bend in the curve of agent adoption around December ‘25 / January ‘26.
• Increased enterprise spending will start showing up in Q1’26 OpEx financial results, but will likely show up as a sharp increase in Q2’26 results.
• Since January, there has been increased pressure at many companies for engineers to “Tokenmaxx”. This is an incredibly easy (and costly) leaderboard to game.
• See, for example:
  • The Pulse: ‘Tokenmaxxing’ as a weird new trend
  • More! More! More! Tech Workers Max Out Their A.I. Use
• “Tokenmaxxing” is fun and memetic, but likely has a short lifespan before you see the median engineer try to game the leaderboard numbers – or at least, change their decision-making on the margin to make less efficient use of tokens.

In the short term, I’d bet that some companies will start imposing soft token budgets on engineers, while others will continue to soft-allow unlimited spend – either for legitimate reasons (more effective/efficient uses are discovered; frontier labs have abundant compute which changes the marginal return calculus), or for memetic/signaling reasons.

There are innumerable ways to make positive-value use of AI. In the short term, Tokenmaxxing – especially the explicit encouragement of Tokenmaxxing in the absence of clear definable output value – is not such a way.