Devin AI Usage Tracking: What Cognition's ACU Analytics Show (and Still Miss)
Devin AI usage tracking spans five settings pages and a per-user ACU cap. Here is what Cognition's own analytics show admins, and where the gaps are.
Pranav Lakhani
CTO, CloudByte AI
TL;DR
- An ACU (Agent Compute Unit) meters Devin's actual session work: planning, context gathering, code and browser actions, and VM runtime. Windows sessions cost about 9% more than Linux; idle sessions stop billing after roughly 0.1 ACU.
- Devin AI usage tracking lives across five surfaces: Enterprise Settings > Consumption, Organization Settings > Consumption Analytics, Session Insights (per-session ACU cost), My Analytics (personal usage), and per-user usage policies that cap monthly ACU spend.
- Devin Review, Cognition's separate PR-review product, is excluded from per-user ACU limits but still draws from the shared enterprise ACU pool, so org-wide spend can rise from review activity even while every individual stays under their cap.
- Self-serve pricing runs Free, Pro ($20/month), Max ($200/month), and Teams ($80/month minimum for up to 200 members, full seats at $40/month). Enterprise is billed per ACU at a rate set in the customer's order form.
- None of Devin's own analytics show how its ACU spend relates to Claude Code, Cursor, or Copilot activity running on the same team, which is the gap a cross-tool tracking layer has to close.
A Devin AI invoice doesn't list hours or seats. It lists ACUs, Cognition's own unit for the planning, browsing, and code execution an agent burns through while it works unattended. Most engineering managers approve a Devin pilot, then spend their first full month figuring out what an ACU buys them.
That question has a real answer. Cognition ships five separate admin surfaces for Devin AI usage tracking, a per-user spend cap, and a documented formula for what counts against it. Here is what each surface shows, where the limits have gaps, and how that compares to tracking a Claude Code or Copilot rollout.
What does Devin AI's admin console track?
Devin AI usage tracking is split across five distinct surfaces, not one dashboard, and each one answers a different question.
| Surface | What it shows | Who sees it |
|---|---|---|
| Enterprise Settings > Consumption | Enterprise-wide ACU totals and billing | Enterprise admins |
| Organization Settings > Consumption Analytics | Org-level ACU totals, with a per-user breakdown | Organization admins |
| Session Insights | The ACU cost of one specific session | Any user with session access |
| My Analytics (Settings > Personal) | A user's own ACU consumption across every org in the enterprise | Users with "View Personal Analytics" permission |
| Usage policies | Per-member monthly ACU caps, set by admins | Enterprise admins configure; members see their own cap |
That split matters for who gets visibility. A line engineering manager with Organization admin rights can see every teammate's ACU totals. A developer without the personal-analytics permission can't even check their own consumption outside of Session Insights, one session at a time. Cognition's docs are explicit that personal analytics is scoped to the user: "users only ever see their own consumption, not anyone else's."
What is a Devin ACU, and how is it different from a seat license?
An ACU prices the actual work a session does, not a flat monthly fee for access.
ACUs accrue from four things, according to Cognition's own documentation:
- Action complexity: planning, context gathering, and the code or browser actions a session takes
- Virtual machine runtime: the time the session's environment stays provisioned and running
- Network bandwidth: typically a small share of total usage
- Platform: Windows sessions run about 9% higher than equivalent work on Linux
Two consumption rules shape what a bill looks like. Sessions stop accruing cost after roughly 0.1 ACU of inactivity, so an agent left idle doesn't quietly run up a tab. And Session Insights buckets every session into a size class: XS at 2 ACUs or less, up through XL at more than 20 (enterprise account thresholds scale ten times higher).
Self-serve plans (Pro, Max, Teams) don't bill in raw ACUs. They use included usage quotas plus on-demand credits, with the same work-based metering underneath. Only Enterprise contracts are billed directly in ACUs at a rate set on the order form.
How much does Devin AI cost, and how does pricing map to usage tracking?
Devin's five plans split cleanly into individual self-serve tiers and a team tier with two different seat types.
| Plan | Price | Who it's for |
|---|---|---|
| Free | $0 | One member, limited included usage |
| Pro | $20/month | One member, daily and weekly quota plus pay-as-you-go credits |
| Max | $200/month | One member, a larger weekly quota with no daily cap |
| Teams | $80/month minimum, up to 200 members | Full seats at $40/month (Pro-equivalent quota) or free Flex seats drawing from shared team credits |
| Enterprise | Custom, billed per ACU | Organization-wide ACU limits, audit controls, SSO |
The Teams plan's Flex seat is the detail worth flagging to a budget owner. A Flex seat costs nothing per month, but every session it runs pulls from the team's shared credit pool. A team that loads up on Flex seats to keep the sticker price low can still burn through shared credits fast if several people run sessions on the same day. The per-member usage picture only gets visible once an admin checks Consumption Analytics, not from the invoice line items alone.
Can admins cap how much ACU a developer burns, and what happens when they hit it?
Yes. Enterprise usage policies let admins set a hard monthly ACU ceiling per member, and that ceiling applies everywhere the member works.
A member's Devin Desktop usage, CLI usage, and cloud session usage all draw against the same single per-user limit. The cap resets monthly, aligned to the contract's billing cycle. Once a member hits it, new work is blocked on every surface, and Cognition's docs confirm the member sees a notification explaining the block, with an option to request more ACUs from their analytics page.
That is a cleaner mechanism than most native AI-tool consoles offer; GitHub Copilot and Claude Code's own telemetry don't expose a comparable per-developer hard stop. The limitation is scope, not enforcement. The cap only governs Devin's own ACU consumption. It has no visibility into whether that same developer is also running expensive agentic sessions in Claude Code or Cursor the same week.
Does Devin Review get tracked separately, and is that a gap?
Devin Review, Cognition's PR-review product, is billed from the same ACU pool as regular sessions but exempted from per-user limits, which creates a real blind spot for anyone watching individual caps.
Devin Review turns large, complex pull requests into organized diffs with explanations, separate from the autonomous coding sessions Devin runs end to end. The two products are distinct, but Cognition's own docs state plainly that Review "consumes ACUs from your enterprise's ACU pool, the same pool used by Devin sessions and other Devin products." The enterprise consumption dashboard breaks that pool down by product, with a dedicated Review line in the daily chart and a per-repository view.
The gap: per-user usage policies explicitly don't count Review usage toward an individual's monthly cap. A team can watch every developer sit comfortably under their personal ACU limit while enterprise-wide spend climbs because Review activity, uncapped at the individual level, keeps drawing from the shared pool. Org admins who only check per-user compliance will miss that pattern. The fix is checking the Review line on the consumption dashboard directly, not inferring it from individual caps.
How does Devin AI usage tracking compare to Claude Code and Copilot?
Devin's ACU model gives richer built-in admin tooling than most IDE assistants, but all three tools stop at their own walls.
| Devin AI (native) | Claude Code (native) | GitHub Copilot (native) | CloudByte AI | |
|---|---|---|---|---|
| Billing unit | ACU (work-based) | Token (model + type) | Per-seat license | N/A, tracks underlying activity |
| Per-developer breakdown | Yes, via Consumption Analytics | No, aggregate only via Anthropic's Analytics API | Partial, via Copilot's own usage metrics | Yes, by design |
| Per-task/per-PR cost | Yes, via Session Insights | No native per-PR view | No per-PR cost signal (flat seat fee) | Yes, for Claude Code sessions |
| Admin spend cap | Yes, per-user ACU policy | No | No | N/A |
| Visibility across other AI tools on the same team | No | No | No | Yes, for Claude Code and the surrounding tool stack |
Devin's admin console is a genuine step up from Claude Code's own telemetry, which reports aggregate token counts per model per day and nothing at the developer or PR level. But it's still a single-vendor view. A manager running Devin alongside Claude Code, Cursor, or Copilot, which most AI-forward teams do, gets five different Devin dashboards and zero combined picture of total AI spend or which seats across every tool sit unused.
What does Devin AI's own analytics still miss for engineering managers?
Devin's analytics answer "how much did Devin cost," not "is our overall AI coding investment working."
Three gaps show up consistently once a team moves past the pilot phase:
- No cross-tool spend view. Devin's consumption dashboards only ever total Devin's own ACUs. A CTO comparing Devin spend against Claude Code token cost or Copilot seat fees has to pull three separate reports and reconcile them by hand.
- No ghost-seat detection outside Devin itself. Devin can tell you a Flex seat is drawing shared credits, but it has no concept of a Claude Code or Copilot seat sitting unused elsewhere in the same org.
- No audit trail linking AI agent work to delivery outcomes. ACU totals say what was spent, not whether the merged PRs it produced actually shipped, shortened cycle time, or needed heavy human rework afterward.
CloudByte AI doesn't close the first gap by ingesting Devin's own session data. Devin runs as a hosted cloud agent with no local SDK or BYOK hook the way Claude Code or OpenAI's API does, so there's no equivalent native integration today. What CloudByte AI does cover is the rest of a mixed-tool team's stack: per-PR token cost and session capture for Claude Code, ghost seat detection across the AI license portfolio, and prompt governance, so a team running Devin for autonomous tasks and Claude Code for everything else isn't flying blind on the half of its AI spend that Devin's own console can't see.
See how CloudByte AI tracks AI coding activity across a team's full tool stack, or check current plans to see where a mixed Devin-plus-Claude-Code rollout fits.
Frequently asked questions
An ACU (Agent Compute Unit) is the unit Devin bills for actual session work: planning, context gathering, code and browser actions, and the virtual machine time a session consumes. Windows sessions run about 9% higher than equivalent Linux work, and a session stops accruing cost after roughly 0.1 ACU of idle time.
Enterprise admins see enterprise-wide totals at Settings > Consumption. Organization admins see org-level totals and per-user breakdowns at Settings > Consumption Analytics. Any user can see the ACU cost of one session in Session Insights, and users with the right permission see their own usage under My Analytics.
Yes. Enterprise admins set a monthly ACU cap per member under usage policies. A member's Devin Desktop, CLI, and cloud session usage all count against that single limit, it resets on the billing cycle, and new work is blocked on every surface once the cap is hit.
No, but it still spends from the shared pool. Devin Review usage is explicitly excluded from per-user ACU limits, yet it draws ACUs from the same enterprise pool as regular Devin sessions and shows as its own line on the consumption dashboard, so org-wide spend can still climb from review activity alone.
Devin meters a single number, ACUs, across five admin surfaces inside its own console. Claude Code's native telemetry reports token counts per model per day with no per-developer or per-PR view. Neither gives a manager a cross-tool picture if the team also runs Claude Code, Cursor, or Copilot alongside Devin.