A new round of data from engineering intelligence firms suggests the AI coding boom may be quietly inflating the metrics it claims to improve. In a TechCrunch report published April 17, Tim Fernholz laid out what he called the "tokenmaxxing" paradox: engineers are shipping more code and burning more tokens, but a growing share of that output is being revised or discarded almost as quickly as it is accepted.
The numbers behind the paradox
Faros AI, analyzing two years of customer data, reported in March that code churn — lines of code deleted relative to lines added — has risen 861% in teams with high AI adoption. GitClear's January 2026 report found that regular AI users averaged 9.4x higher churn than non-AI counterparts. Jellyfish, which looked at 7,548 engineers across Q1 2026, found that engineers with the largest token budgets produced the most pull requests, but delivered "two times the throughput at ten times the cost of tokens."
Waydev, an engineering intelligence vendor founded in 2017 that now works with more than 50 customers covering over 10,000 engineers, added another data point: initial AI code acceptance rates sit at 80–90%, but real-world rates fall to 10–30% once revisions are counted. "This is a new era of software development, and you have to adapt," Waydev CEO Alex Circei told TechCrunch.
Why "tokens per engineer" became a KPI
As Claude Code, Cursor, and Codex rolled out across enterprise fleets, many engineering orgs started tracking token consumption as a proxy for AI adoption. The problem, according to the reports cited by TechCrunch, is that high token spend correlates with more pull requests on the surface but also with more rework underneath — the same feature cycling through accept, revise, and rewrite in quick succession.
Reid Hoffman, weighing in on the debate earlier this week, said token tracking can be useful for gauging adoption but warned against treating it as a direct productivity metric without richer context.
Implications for engineering leaders
The findings will sharpen an already loud debate inside CTO offices about whether AI coding spend is pulling its weight. Atlassian's recent $1 billion acquisition of engineering intelligence startup DX underscored how much appetite exists for better measurement tools, and the Faros, Jellyfish, GitClear, and Waydev numbers suggest that the old "lines of code" trap has simply been re-skinned as a "tokens consumed" trap.
For developers, the implication is more pointed. If 861% more code churn is the real cost of today's AI-assisted workflows, the productivity story looks less like a step-function leap and more like a very expensive treadmill — one that may only pay off for teams willing to pair token budgets with disciplined review, testing, and churn metrics of their own.
