Cognition Fights Token Excess with AI Efficiency Grading System

Cognition, the startup behind autonomous coding agent Devin, is pushing back against corporate AI "tokenmaxxing" with an internal system that grades engineers on productivity per dollar spent—and rewards the most efficient with higher spending limits. The efficiency metric divides the human-equivalent engineering hours produced by each AI session by its token and infrastructure costs. Cognition validated these estimates against feedback from 126 users across 258 sessions, achieving a 0.74 log correlation with real-world outcomes on held-out data. Engineers who meet productivity quotas receive dynamically increased spending limits, inverting the traditional model where heavy token consumption was treated as a mark of power-user status.