Cognition CEO recommends replacing “token volume” manipulation and volume-farming performance assessments with an evaluation based on actual output.

According to Beating monitoring, Scott Wu, co-founder and CEO of AI programming startup Cognition, said on the Founders podcast that some companies have gone too far by assessing programmers based on the number of AI Tokens employees consume; the standard should be the actual completed tickets and output.

This trend is known in Silicon Valley as tokenmaxxing. To gain an advantage in internal AI usage dashboards or performance evaluations, employees reflexively make heavy calls to AI programming tools such as Claude, Codex, or Cursor. Jacob Lauritzen, technical director of the legal AI startup Legora, also criticized this on a podcast this month, saying that many people display Token usage in performance reviews, which causes employees to inflate usage purely to make the data look good—a foolish way to evaluate.

Andrew Feldman, CEO of chip developer Cerebras Systems, said at a Bloomberg conference this month that providing employees with unlimited Token allowances was foolish from the very beginning. Feldman advised that companies should place greater emphasis on cost-effectiveness when using AI tools—using low-cost open-source models rather than abusing top-tier closed-source models for ordinary tasks.