$1M Unlimited Token Test: Blindly Delegating to AI is Costlier and More Exhausting

dotey · x · 2026-08-11

An experiment with a $1 million budget providing unlimited AI tokens to a 20-person team revealed unexpected outcomes: AI costs can exceed human labor, and the process actually leaves workers more fatigued.

The most significant productivity gains occurred during the organization's initial 0-to-1 adoption of AI. However, because corporate structures are designed for humans, efficiency hits a hard ceiling even with infinite compute. Critically, over-reliance on AI causes employees to outsource their thinking, diminishing their grasp of project details.

Developers argue that merely boasting about token consumption is a toxic remnant of the 'Token Max' era. The optimal approach is strategic allocation: routing complex architecture to premium models, moderate tasks to open-source alternatives (like GLM or DeepSeek), and simple queries to local models with caching. This tiered strategy yields a significantly higher ROI than indiscriminately dumping everything into top-tier APIs.

Related event: $1M Experiment Shows Unlimited AI Tokens Unsustainable(2 posts)→

Original post →

More from coding & agent

coding & agent channel →