Skip to Content
IDEsDevin

How to decrease Devin spend

Devin  is Cognition’s autonomous AI software engineer — cloud VM sessions, Devin CLI, and Devin Desktop (Cascade, Devin Local). Each surface meters differently, but spend spikes from the same patterns: vague cloud sessions, oversized rules, idle MCP schemas, marathon threads, and frontier models on routine work.

Quick answer: default to Adaptive or Devin Local, trim always_on rules and unused MCP, scope cloud tasks with acceptance criteria, and start a new conversation per task. For the general stack see Where to start.

Where do I see Devin usage? Settings → Plans  for quota and on-demand credits; Plan Info on Desktop for token groups. Full breakdown: Devin token usage.

Quick checklist

  1. Cloud: scope tasks with acceptance criteria — Cloud ACUs and Instructing Devin.
  2. Desktop: open Devin Settings → Plan Info — see Devin token usage for ACUs, quota, and credits.
  3. Default to Adaptive or Devin Local (see Devin Local vs Cascade); not frontier on every prompt.
  4. Switch always_on rules to glob or model_decision.
  5. Disable MCP you are not using. Stay under the 100-tool cap on Desktop.
  6. CLI: use local Devin for quick fixes; /handoff longer work to cloud — CLI guide.
  7. New conversation per task; review Session Insights after heavy cloud sessions.

Typical impact: 30–60% switching Cascade → Devin Local; 20–50% trimming rules and MCP; meaningful ACU cuts on scoped cloud tasks with clear success criteria.

Three surfaces, three meters

SurfaceWhat it isMeterDeep dive
Cloud DevinAutonomous sessions in a VM (web, Slack, Linear, API)ACUs, quota, or on-demand creditsCloud ACUs
Devin CLILocal terminal agent; /handoff to cloudShares plan quota / ACUsCLI
Devin DesktopIDE with Cascade, Devin Local, Tab, CommandQuota, credits, or ACUs in Plan InfoToken usage, Devin Local

Official docs describe Devin as handling most tasks a human engineer could finish in about three hours — with higher success when prompts include explicit completion criteria and verifiable checks (CI, tests, browser QA).

How Devin bills (overview)

Self-serve plans (Free, Pro 20/mo,∗∗Max∗∗20/mo, **Max** 200/mo, Teams $80/mo minimum) share a quota across cloud sessions, CLI, and Desktop. Past quota, on-demand credits keep you working; credits roll over month to month. Enterprise contracts bill in ACUs (Agent Compute Units).

Plan typeUnitsWhere to read
Self-serveDaily/weekly quota + on-demand creditsSettings → Plans 
EnterpriseACUs per order formConsumption 
Desktop detailToken groups, credits, ACUsToken usage

Cloud sessions sleep when idle — no usage while sleeping, waiting on you, or while tests run. See Usage .

Adaptive  routes simple desktop and CLI work to lighter models. Enterprise: Adaptive is off by default until admins enable it.

Guides

GuideWhat you’ll learn
Devin token usageACUs, quota, on-demand credits, Plan Info, enterprise APIs
Cloud ACUsSession scoping, Session Insights, sleep, Linux vs Windows
Instructing DevinPrompts, success criteria, session size, pre-task checklist
Devin CLIInstall, local work, /handoff to cloud, Adaptive
Devin Local vs CascadeLocal agent, Cascade migration, “Cascade is disabled” troubleshooting
MCPCloud marketplace vs Desktop schemas, 100-tool cap
Rulesalways_on vs glob, global rules, AGENTS.md, Knowledge

Model routing

Default cheap, escalate on failure — Model routing:

TierChoice
RoutineAdaptive or Devin Local
Multi-fileSWE-1.x / mid-tier Sonnet
Deep debugFrontier / thinking — last resort
Autonomous backlogCloud Devin with scoped milestones

Anti-patterns

Anti-patternWhy it fails
Cloud Devin for laptop-sized fixesVM + ACU overhead — use CLI or Devin Local
Arena mode for routine editsParallel agents bill independently
always_on rules for style guidesInput bloat every message
Marathon thread across tasksCache read compounds
Vague “improve the app” cloud promptExploration loops — see Instructing Devin
Every MCP in the marketplace enabledTool schema and action overhead
Last updated on
MCPToken usage