Cognition Links AI Spending Limits to Engineering Productivity

What changed

Cognition is promoting a framework that replaces flat AI spending caps with usage limits tied to efficiency. It says automatic cap increases of up to 22 times helped one engineer raise merged pull requests from about 30 to about 80 a week.

The proposed measures include productive hours per unit of spend, plus reviews of vague prompts, costly model modes and mixed-up sessions through Devin Coach.

Why it matters

A blanket $20, $50 or $100 monthly limit treats every engineer as if they use AI in the same way. Cognition’s argument is more practical: spending should rise when it produces useful work, and tighten when it produces waste.

That shifts the manager’s job from policing a bill to inspecting how the tools are used. The interesting part is not the larger allowance itself. It is the attempt to connect AI costs to engineering output.

The productivity example is still a claim, not an independently verified result. Cognition does not identify the engineer, company, measurement period or baseline conditions. That makes the 30-to-80 figure a useful prompt for testing the framework, but a weak basis for copying its limits wholesale.