Tokenmaxxing loses favor as companies scrutinize AI costs

Corporate enthusiasm for “tokenmaxxing,” the practice of maximizing token use in generative AI tools, is fading as companies see higher bills without comparable productivity gains. Moody’s Ratings’ Vincent Gusdorf said workplaces are realizing AI can easily create unnecessary work and that new tools need to be used more carefully.

Silicon Valley leaders had recently treated heavy usage as a sign of ambition. Sam Altman said in May he was excited about tokenmaxxing startups, Jensen Huang said “if your $500K engineer isn’t burning $250K in tokens, something is wrong,” and Meta ran an internal competition rewarding token usage. The trend lifted revenue for major large language model providers but drew criticism from executives including Satya Nadella and Alex Karp, who questioned the value and data tradeoffs for enterprise customers.

Businesses are now looking more closely at returns and turning to model routing, where simpler tasks go to cheaper systems and harder work goes to more capable models. Bain & Company’s Jue Wang said token costs for some clients have been “doubling, almost every other month,” while cheaper open-source models from Chinese startups are giving developers lower-cost options.