The defining trend of 2026 in software engineering is “tokenmaxxing” — burning as many AI tokens as possible as a proxy for productivity. The data suggests it is making code worse, and developers less capable without it.
In May, TechCrunch reported that developers are increasingly refusing to take on coding tasks without access to AI assistants. Some junior engineers describe themselves as “unable to function” without Copilot or Claude Code. But a growing body of research — from Anthropic, Faros AI, and Jellyfish — suggests that this dependency comes with real costs.
What’s new. The Faros AI Engineering Report 2026 analyzed telemetry from 22,000 developers across 4,000 teams. The headline numbers look good: epic completion is up 66%, task throughput is up 33.7%. But the quality metrics tell a different story. Bugs per developer are up 54%. Incidents per pull request have increased 242%. Code churn under high AI adoption is up 861%.
Faros calls this the “Acceleration Whiplash” paradox: teams ship more code faster, but the code is measurably worse, and the debugging burden compounds over time.
Anthropic’s own research, published in January 2026, found that developers using AI scored 50% on code comprehension tests, compared to 67% for unassisted coders. AI-assisted developers were faster at writing code but significantly worse at reading, understanding, and debugging it — skills that matter when things go wrong.
The key angle. The term “tokenmaxxing” emerged in early 2026 to describe engineers who measure their productivity by how many AI tokens they consume. The New York Times documented the phenomenon in March, reporting that the practice had turned into a status game: more tokens equals more output, or so the thinking goes.
Meta reportedly maintained an internal tokenmaxxing leaderboard, which was shut down in April after leaking to the press. Jellyfish’s data showed that extreme token use delivers diminishing returns — burning more AI tokens does not proportionally improve output, and the correlation between token consumption and shipped value flattens sharply beyond a modest threshold.
The worst-case scenario is junior developers who have never written code without AI assistance. They produce code faster but cannot debug it, cannot review it critically, and cannot function when the AI goes down. As one engineering manager told TechCrunch: “They’re not learning how to program. They’re learning how to prompt.”
Context / What’s next. The backlash against tokenmaxxing is building. Companies that adopted AI coding tools aggressively are starting to measure outcomes rather than token counts. The LeadDev conference in April 2026 featured multiple sessions on meaningful AI productivity metrics, and several major tech firms are reportedly revising their AI adoption strategies.
But the genie is not going back in the bottle. Developers who have experienced AI-assisted coding are unlikely to give it up voluntarily. The question is whether the industry can develop training and measurement practices that preserve the throughput gains without sacrificing the foundational skills that make good engineers.
The big picture. The AI coding assistant market is in a strange place: the tools are genuinely useful, widely adopted, and getting better every quarter. But the early evidence suggests they are also creating a generation of developers who are faster at producing code and worse at understanding it. That tradeoff may be acceptable for senior engineers who already have deep foundations. For juniors learning their craft, it is a risk the industry has not fully reckoned with.
Sources: TechCrunch (May 29, 2026); TechCrunch (April 17, 2026); Anthropic (January 2026); Faros AI (March 2026); New York Times (March 20, 2026); Business Insider (May 2026); LeadDev (April 27, 2026)