High AI token usage doubles developer output at ten times the cost

Developers are increasingly embracing "tokenmaxxing"—the practice of maximizing AI token usage in coding tools like Claude Code, Cursor, and Codex—but new data reveals it's not boosting true productivity as much as it seems. Engineers with the highest token budgets generate far more code pull requests, yet at a dramatically higher cost and with frequent revisions that erode long-term gains, according to analysis from developer productivity platforms.

Token budgets represent the amount of AI processing power a developer can consume, and in Silicon Valley, high usage has become a status symbol, much like conspicuous consumption. As reported by TechCrunch, managers initially see impressive code acceptance rates of 80% to 90% for AI-generated suggestions, but real-world rates drop to 10% to 30% after engineers spend weeks revising and rewriting the output. This churn means more code volume upfront, but less efficiency overall, undercutting claims that AI tools are revolutionizing software development.

Jellyfish, an intelligence platform tracking AI-integrated engineering, examined data from 7,548 engineers across companies in the first quarter of 2026. They found that top token users—those in the 90th percentile consuming around 380 million tokens per month, compared to the median of 51 million—produced up to 2.15 pull requests per week, double the 0.77 from low users. However, this came at 10 times the token cost, with high-end developers using 69 million tokens per pull request versus 7 million for typical ones. The result: doubled throughput without proportional value, as the firm detailed in its blog.

This trend highlights a broader gap in AI adoption, where insiders chase metrics like token spend while overlooking outputs that matter, such as sustainable code quality. Companies like Meta, Microsoft, and Salesforce are seeing developers inflate AI usage to meet internal targets, treating tokens as a badge of honor rather than a tool for efficiency. Engineering leaders, per insights from The Pragmatic Engineer newsletter, are grappling with how to measure productivity beyond lines of code or raw AI consumption, especially as subsidies for these tools wane—Anthropic has ended enterprise plans, and Uber exhausted its 2026 budget in just three months.

The implications extend to businesses footing the bill, as tokenmaxxing drives up expenses without scalable productivity. Affected parties include engineering teams facing revision overload, managers misled by surface-level metrics, and companies questioning AI investments. Looking ahead, per-engineer AI budgets are expected to proliferate, pushing a shift from maximizing tokens to optimizing them for real value, as Jellyfish advises. Without this pivot, the hype around AI coding agents risks stalling genuine progress in developer workflows.