Skip to Content
IDEsDevinCloud ACUs

Devin Cloud ACUs

Devin Cloud is the autonomous agent that runs in Cognition’s workspace — shell, embedded IDE, and browser — to plan, implement, test, and open PRs. It bills through your plan’s quota and on-demand credits (self-serve) or ACUs (enterprise), not just raw tokens.

Hub: Tokenminning in Devin.

What usage includes

Per official usage docs :

ComponentNotes
LLM inferencePlanning, edits, review loops
ActionsShell, browser, tool execution
VM timeUsually small fraction of total
Windows sessions~9% premium vs Linux (Ubuntu)

Usage accrues from work Devin actually performs — not from idle waiting.

Sleep and idle behavior

Cloud Devin sleeps when idle. No usage while:

  • Sleeping
  • Waiting for your reply
  • Waiting for a test suite to run
  • Setting up and cloning repositories (aside from minimal VM keep-alive)

Devin typically sleeps after roughly 0.1 ACUs (or the equivalent quota / on-demand credit) of inactivity. Awake-but-idle time generally adds up to very little.

Measure cloud spend

ToolUse
Session InsightsPer-session ACU/quota consumption, issue timeline, improved prompts
Settings → Plans  (self-serve)Quota remaining, on-demand balance
Settings → Consumption (enterprise)Org rollups
Session metrics API Category breakdown

Desktop metering: Token usage.

Session Insights

Session Insights  analyzes completed sessions at no extra cost. Use it to find ACU waste patterns.

Overview metrics:

MetricWhat it tells you
ACU usageTotal compute for the session — compare similar tasks
User messagesHigh count → underspecified initial prompt
Session sizeXS / S / M / L / XL composite — L and XL are unhealthy
CategoryHow Devin classified the work (bug fix, feature, refactor, etc.)

Session size thresholds (self-serve ACU scale; enterprise scales ACU thresholds ×10):

SizeACU thresholdUser messages
XS≤ 2≤ 2
S≤ 5≤ 5
M≤ 10≤ 10
L≤ 20≤ 20
XL> 20> 20

Overall size is the larger of the ACU-based and message-based classifications.

Tabs to review after expensive sessions:

  • Issue Timeline — build failures, environment gaps, retry loops
  • Actionable Feedback — rewritten prompt with highlighted improvements; machine/repo action items
  • Knowledge Usage — which Knowledge  items helped vs misled Devin

Common patterns:

PatternLikely causeFix
High ACU, few user messagesAutonomous struggle, retriesEnvironment setup, clearer scope — Instructing Devin
Many messages, low ACUConstant course correctionsFront-load context; add Knowledge
L/XL session sizeTask too broadSplit into focused sessions
Misleading Knowledge flaggedStale org instructionsUpdate or delete Knowledge items

Scope tasks to cut ACUs

Official guidance: delegate clearly scoped tasks with a well-defined end goal; keep prompts and sessions short; avoid many different tasks in one session; split big projects across sessions (no concurrent session limits on most plans).

PatternACU impact
Vague “build me an app”Long exploration loops
One message, no acceptance criteriaBack-and-forth clarification
Local-sized fix in cloudVM overhead — use CLI or Devin Local
Open-ended “stand by” threadKeeps session context hot

First message checklist (from instructing guidelines ):

  • Defined end goal and explicit success criteria (tests, CI, browser checks)
  • Files or areas in scope
  • Examples, links, or patterns to follow
  • Milestones for large work — one session per milestone

Pre-task checklist highlights:

  1. Clear start, end, and verifiable success criteria
  2. Examples, filenames, docs links, or MCP connections for context
  3. Test/lint/CI validation Devin can run autonomously
  4. Session kept to XS–M size where possible

vs Devin Desktop and CLI

WorkloadPrefer
IDE pair programmingDevin Local / Adaptive in Desktop
Quick terminal fixDevin CLI
Autonomous multi-hour projectCloud with phased sessions
Hand off from localCLI /handoff — packages context and branch

Install CLI: curl -fsSL https://cli.devin.ai/install.sh | bash

When cloud Devin excels

From When to use Devin :

  • Parallel backlog work (Linear/Jira tickets, bug reports)
  • Migrations and refactors with verifiable checkpoints
  • PR review with Auto-Fix closing the loop on CI failures
  • Scheduled sessions for recurring triage
  • Tasks you can finish in ~3 hours with clear success criteria

Use Ask Devin to scope before burning a full implementation session.

Guardrails

Last updated on