In 2026, AI has become so embedded in software development that many programmers refuse to code without it. A failed replication study by the AI safety lab METR, combined with the rise and fall of "tokenmaxxing" and mounting evidence of long-term maintenance burdens, reveals the growing hidden costs of this dependency — even as companies like Cognition push AI coding agents as collaborative tools.
METR’s failed experiment tested the willingness of developers to work without AI assistants. The lab’s original 2025 study had found that while developers felt more productive with AI, overall task times increased due to correcting AI errors, guiding the tool, and waiting for responses. When METR tried to replicate the research in 2026, no one would join the control group that required working without AI. Instead, a survey was published where respondents perceived themselves as twice as valuable — an assessment at odds with objective data.
Tokenmaxxing, the practice of using AI token consumption as a productivity proxy, became a short-lived trend. Amazon shut down its internal leaderboard, Kirorank, after employees gamed it by running AI agents excessively with no meaningful output gain. Uber burned through its entire 2026 AI budget in just four months, and COO Andrew Macdonald admitted on a podcast that the spending did not increase measurable productivity.
The maintenance cost trap is now backed by multiple sources. Programmer James Shore warned in a widely shared blog post that doubling coding speed without halving maintenance costs creates permanent indenture. Independent research from Singapore Management University pointed to long-term liabilities introduced by AI-generated code. Code Rabbit’s analysis of open source pull requests found AI-produced code caused 1.7 times more problems than human-written code. Even the CEO of reliability startup Entelligence AI claimed companies spend 44% of AI tokens fixing bugs created by AI itself.
Amid these concerns, Cognition, maker of the AI coding agent Devin, raised $1 billion at a $26 billion valuation. CEO Scott Wu insists Devin is a “buddy” that helps programmers, not a replacement. Wu, a lifelong coder, describes Devin as performing at a junior to mid-level engineer level, handling tedious tasks like updating legacy software. Strikingly, 89% of code committed by Cognition’s own engineers is done by Devin, yet Wu stresses that humans must retain responsibility for architecture and security.
The coding profession faces an inflection point. The evidence suggests that while AI assistants are here to stay, disciplined practices — understanding AI’s strengths, building robust QA for AI-generated output, and treating it as junior work — are essential to prevent spiraling costs and fragile codebases.
