Open Devin
Lesson 7 · Reading · 4 min

ACU & cost governance

ACUs meter the work Devin actually performs: actions taken, VM time, bandwidth. The corollary matters as much as the rule: there is no consumption while a session sleeps, waits for the user's reply, waits on external CI, or clones a repo. Actively running tests does consume, because that's real work on the VM. Sessions auto-sleep after a short period of inactivity, and Windows sessions carry a modest surcharge over Linux. Exact figures live in the docs and change over time, so quote the docs, not this course.

What drives a session's cost is mostly what drives its length: task complexity, prompt specificity, session length, and how much back-and-forth it takes. The biggest lever here is context quality, and that was Cloud 102; this lesson is about governing cost, not re-teaching prompting.

Your governance toolkit, from broadest to narrowest:

  • Enterprise consumption dashboard (Settings → Consumption): total usage across the account.
  • Per-org consumption analytics: which orgs consume what, on the billing boundaries from Lesson 1.
  • Per-org ACU limits: a hard stop. When an org hits its limit, all Devin activity in that org halts and users are told to contact the enterprise admin. Nothing throttles gracefully past the line, so set limits with headroom.
  • Session Insights: per-session cost and behavior, for diagnosing individual outliers.
  • Consumption and metrics APIs: export everything into your own BI stack.

Session Insights turns cost anomalies into coaching. Two patterns cover most outliers. High ACU with few user messages means the task was too large or vague for one session; coach the user to scope smaller and front-load a stronger brief. Many messages with low ACU means drip-feeding: the user is steering step by step instead of delegating; coach them to write the full brief up front and walk away.

Put together, the admin's cost loop is: detect the outlier on the dashboards, diagnose it with Session Insights, and fix the systemic cause, whether that's a slow environment, a missing knowledge item, or a prompting habit. You're not policing individual sessions; you're fixing the things that make many sessions expensive. Limits are the backstop for the month something goes wrong; the dashboards are where the ongoing steering happens.

Back to course← Previous lesson
© 2026 Cognition AI, Inc.Made with by the Cognition team