Two years into the widespread adoption of AI coding tools, a new phenomenon is surfacing: some teams are seeing their code commit counts double, yet product delivery hasn't accelerated, and bugs are on the rise.
TechCrunch recently reported on a term: Tokenmaxxing. It describes developers who treat their AI tool token budgets as a status symbol—the larger the token quota, the more productive they believe themselves to be. But the data tells a different story.
More Tokens, Faster Code Deletion
Consider the numbers:
- GitClear, January 2026: Developers using AI tools have a code churn (deletions and modifications relative to additions) that is 9.4 times higher than their non-AI-using colleagues.
- Faros AI, March 2026: Analyzing two years of customer data, high AI adoption was linked to an 861% increase in code churn.
- Jellyfish, Q1 2026: Engineers with the largest token budgets had twice the throughput of average developers, but at a cost 10 times higher.
This isn't to say AI tools are useless. Rather, much of the code counted as "productivity gains" often doesn't survive more than a few weeks.
The Truth Behind the 80% Acceptance Rate
Engineers accept 80-90% of AI-generated code on average. This high number is often used by companies as a core metric for "AI ROI."
But Alex Circei, CEO of Waydev, tracked data from over 50 companies and more than 10,000 engineers, and found a pattern:
The initial acceptance rate is 80-90%, but when you look again a few weeks later, only 10-30% of that code actually remains.
In other words, when an engineer hits "accept," it doesn't mean the code is truly useful. A large amount of AI-generated code is quietly deleted or rewritten in subsequent iterations. These deletions are never counted in reports of "how much time AI saved me."
When Token Budgets Become KPIs
The emergence of the term "Tokenmaxxing" signals that AI tool adoption has shifted from a technical problem to an organizational management issue.
When engineering teams start treating "monthly AI token consumption" as a performance metric, problems arise. Engineers begin actively using AI to generate large volumes of code to demonstrate workload, rather than carefully judging which code should actually be written by AI.
Atlassian spent $1 billion acquiring the developer analytics platform DX precisely because it saw this trend—companies need tools to understand the true ROI of their AI investments, rather than deluding themselves with token consumption numbers.
The Productivity Bottleneck Has Shifted
AI coding tools are indeed useful. A McKinsey study from February 2026 showed that AI tools compress the time for routine coding tasks by an average of 46%.
But what's compressed is the time to write code, not the time to modify and delete it.
Jellyfish data shows engineers spend 11.4 hours per week reviewing AI-generated code, compared to only 9.8 hours writing new code—a ratio that was reversed in 2024.
Code generation has become faster, but code review has become the new bottleneck. Treating speed as quality is the core mistake of Tokenmaxxing.
Source: CocoLoop, 'Tokenmaxxing' is making developers less productive than they think (TechCrunch)