Cognition, the startup behind the autonomous coding agent Devin, built an internal system that grades engineers on how efficiently they use AI. The best performers then get higher spending limits.
How the grading actually works
Cognition converts each completed Devin session into an estimate of equivalent human engineering hours. Those estimates were validated against feedback from 126 users across 258 sessions. On held-out data, meaning sessions the model had not been trained on, the estimates showed a log correlation of 0.74 with what users reported.
The efficiency score comes from basic division. Take the human-equivalent hours a session produced, then divide by what that session cost in tokens and spend.
Spending limits that work like credit lines
Engineers who hit defined productivity quotas see their spending limits rise dynamically. That flips the logic of tokenmaxxing on its head. Under the old approach, heavy token use was the signal of a power user. Under Cognition’s model, heavy token use with little to show for it is a red flag.
The system went live with customers in mid-2026. It was publicly detailed in reports published on October 15 and 16, 2026.
Alongside the grading system, Cognition released Devin Coach, an AI assistant designed to help engineers write better prompts and use the agent more efficiently. One example of its guidance: advising users on how to break large tasks into smaller segments.
Why a usage-billed company wants you to spend smarter
Cognition bills on usage, so in the short term, more tokens consumed means more revenue. The grading system also builds on an earlier promise. In June 2026, Cognition introduced its AI Productivity Guarantee, which offers up to $10 million in credits if customers do not get value equivalent to what they spend.
Tokenmaxxing drew criticism at firms including Meta, Uber, and Amazon. Cognition’s pivot reflects a broader rethink across the industry about what AI productivity metrics should actually reward.
What this means for AI buyers and the market
For companies paying for AI coding tools, a CFO asking “what did we get for this spend” now has a vendor offering a number in hours rather than a shrug and a token count.
There are open questions worth watching. A 0.74 log correlation leaves room for disagreement on individual sessions, and engineers whose spending limits hinge on the score will notice any misfires. Any metric that controls resources also invites gaming. Goodhart’s law has a long track record in software teams, and tokenmaxxing itself was arguably an example of it.
Cognition is betting that customers would rather pay for results than for volume, and it has put up to $10 million in credits behind that bet.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.