“Tokenmaxxing” is making developers less productive than they think

In the world of software development, the mantra "what you measure matters" rings true. For decades, engineers have debated the best measures of productivity, with lines of code often at the center of discussion. However, as AI coding tools proliferate, determining what metrics should truly reflect developer productivity has become increasingly complex. The rise of token budgets—essentially the AI processing power developers can utilize—has become a point of pride among engineers in Silicon Valley. Yet this focus on input reveals a fundamental flaw in assessing productivity; outcomes should be the priority, not merely the volume of resources consumed. While promoting AI adoption can benefit from such metrics, they may not contribute to actual efficiency. Recent insights from companies specializing in developer productivity analytics are shedding light on this issue. Tools like Claude Code, Cursor, and Codex are enabling developers to produce significantly more accepted code. However, this increase is accompanied by a troubling trend: developers are frequently needing to revise that code, undermining claims of enhanced productivity. Alex Circei, CEO of Waydev, emphasizes this disconnect. His firm, which serves over 10,000 engineers across 50 clients, is tracking these dynamics closely. While engineering managers report high code acceptance rates of 80% to 90%, they often overlook the substantial revisions that follow, which can reduce the effective acceptance rate to between 10% and 30%. In response to the evolving landscape of AI coding tools, Waydev has revamped its platform to provide analytics on the metadata produced by these AI agents, helping managers gain insights into both the adoption and effectiveness of AI in their teams. Despite the potential benefits that analytics companies may highlight, many large organizations are still grappling with the efficient utilization of AI tools. This trend hasn’t gone unnoticed in the industry. Atlassian’s acquisition of DX for $1 billion last year illustrates a growing awareness among major players about the need for better understanding the return on investment from coding agents. Data from multiple sources indicates a consistent narrative: while more code is being produced, a significant portion is not being retained or utilized effectively. Reports from GitClear reveal that although AI tools can boost productivity, regular users experience a staggering 9.4 times higher code churn compared to those who do not use AI. Similarly, Faros AI found an 861% increase in code churn among users heavily reliant on AI over two years. Moreover, Jellyfish's analysis of over 7,500 engineers in early 2026 indicated that while larger token budgets correlated with increased pull requests, the cost of achieving this throughput was disproportionately high. Developers are echoing these statistics in their experiences. Many are overwhelmed by the accumulating technical debt, despite enjoying the capabilities that new tools offer. A notable disparity has emerged between senior and junior engineers, with less experienced developers often accepting a higher volume of AI-generated code and facing greater rewriting challenges. Yet, the sentiment among developers is clear: there’s no turning back. Circei notes, “This is a new era of software development, and adaptation is not optional.” As organizations navigate this transformative phase, they will need to rethink their strategies to harness the full potential of AI in development effectively.