‘Tokenmaxxing’ is making developers less productive than they think

In software engineering, measuring productivity has always been tricky. While the use of AI coding agents has led to a dramatic increase in generated code, a new trend known as 'tokenmaxxing'—the pursuit of higher AI usage quotas—may not be as beneficial as it seems.

Why More Tokens Doesn’t Mean More Productivity

Traditionally, productivity was tracked by outputs like lines of code. Today, the focus has shifted as developers gain bragging rights for using larger AI token budgets. However, token usage simply measures input, not the actual value produced. It can encourage more use of AI tools, but doesn’t guarantee better results or efficiency.

What the Data Reveals

Emerging analytics companies such as Waydev, GitClear, Faros AI, and Jellyfish are closely monitoring the impact of AI tools. Their research shows an increased volume of accepted AI-generated code, but teams also have to spend much more time revising this code. For instance, while initial acceptance rates for AI code can reach up to 90%, real-world retention after necessary revisions drops to between 10% and 30%.

• Waydev highlights how frequent code revisions negate the benefits of high acceptance rates.
• GitClear found that engineers using AI experience 9.4 times higher code churn than those who don’t—greater than the productivity boost itself.
• Faros AI’s data indicates an 861% increase in code churn with heavy AI adoption.
• Jellyfish reports that developers with the largest token budgets produced more code changes but at much higher costs, with only marginal productivity gains.

The Challenge for Modern Engineering Teams

Despite these setbacks, most organizations see AI tools as the future and are not about to give them up. The focus now is on adapting software development processes to leverage AI effectively while minimizing technical debt and unnecessary revisions.

Senior vs. Junior Developers

Studies also point to differences based on experience. Junior engineers are more likely to accept AI-generated code wholesale, often leading to more rewriting, while senior developers exercise greater caution.

As companies learn to navigate this new era, the message is clear: True productivity isn’t measured by AI usage but by the long-term value and maintainability of software.

For more details, read the original article by Tim Fernholz at TechCrunch.