Tokenmaxxing is the practice of maximizing consumption of tokens, the units of text an AI system processes, and treating the volume itself as evidence of productivity rather than measuring the value of the work produced. The term spread through Silicon Valley in early 2026 after Kevin Roose, a New York Times columnist, reported on internal leaderboards in Meta and OpenAI that ranked employees by token use.
The pattern shaped employee behavior at a large scale before any company confirmed the extra usage produced value. On OpenAI’s leaderboard, an engineer processed 210 billion tokens in a single week. While at Meta, an employee built an internal ranking called Claudeonomics that tracked more than 85,000 workers, using a combined 60.2 trillion tokens in thirty days before Meta took the ranking down amid public criticism.
Not every company reached the same conclusion. Farhan Thawar, Shopify's head of engineering, once defended heavy spending as evidence a developer has "an agent workforce underneath them," but Shopify later renamed its leaderboard a usage dashboard, dropping the competitive framing. Amazon reached a harsher verdict with its own leaderboard, Kirorank, after staff began assigning agents needless tasks purely to climb the rankings, driving up computing costs with no matching gain in finished work. Dave Treadwell, an Amazon senior vice president, announced the shutdown by telling staff not to use AI just for the sake of using it, an admission that the metric itself, not just its abuse, was the flaw.
In work automation, tokenmaxxing works against how autonomous agents perform. More tokens do not reliably improve an agent’s output, and pushing past that limit can make the work worse. Data gathered from roughly 100,000 coding-agent sessions found that past 50,000-100,000 working tokens, agents produce more repeated tool calls, more premature terminations, and more fabricated results. Practitioners tie the decline to a buildup of low-value Context crowding out the instructions that matter, not to the window running out of space, so the model has more to weigh and less signal to weigh it against. An agent kept running just to raise a leaderboard score reproduces this failure: it fills its own context with redundant steps, buying visible activity rather than a reliable result, so someone still has to check the work by hand. Shopify's fix is a circuit breaker that kills an agent's access the instant its spend spikes, catching what the company calls runaway agents before the cost compounds.
