Devin pricing: plans, quotas, ACUs, and the real cost

The current Devin price list

The short answer is Free, $20, $200, or a team plan with an $80 floor. That is materially different from the legacy Devin pricing still quoted in search results, where a Core plan bought Agent Compute Units and a public Teams plan cost $500 a month. Those figures describe an earlier billing system, not a new self-serve purchase in August 2026.

The plans cover a broader product line than the original cloud agent. Cognition now presents Devin Cloud, the Devin CLI, Devin Desktop, and Devin Review as parts of one account and commercial system. Devin Desktop is the editor formerly called Windsurf. That consolidation is why a $20 Pro seat cannot be read as merely an editor price or merely an autonomous-agent price. It is entry to Cognition's current product line, subject to its plan quota.

What each plan actually changes

Free is for proving that the delegation model fits your repository. It is not a representative capacity test for a team rollout, because the quota is deliberately light and model availability is limited. It is still enough to run a bounded task, inspect the environment and handoff, and decide whether reviewing an asynchronous result feels better or worse than steering an agent in a terminal. A free trial that never reaches a real pull request tells you almost nothing.

Pro is the ordinary individual plan. The important phrase is daily and weekly quota. A developer can exhaust a short-term allowance during an intense day even if the weekly pool is not gone, while a week of repeated large runs can exhaust the wider allowance. Cognition does not reduce this to a public fixed number of tickets because task consumption depends on model, context, retries, and agent behavior. Treat Pro as a $20 access tier with a variable amount of completed engineering work, not as a bundle of N pull requests.

Max is not simply Pro multiplied by ten. It costs ten times as much, but its practical distinction is that it has a weekly allowance without a daily cap. That matters for lumpy work: a migration weekend, a large backlog batch, or several parallel sessions on one day. It matters less for a developer who uses a small amount every weekday. The decision is therefore about the shape of demand, not only its monthly total.

How quota and overage work

Current self-serve usage is model-weighted. A session consumes tokens and the selected model affects how heavily that work counts. That is more legible than pretending every autonomous minute is equal, but it is not a stable conversion from prompt to finished task. A precise dependency update with a clean test can consume less than an exploratory bug even when the latter changes five lines. Context collection, failed attempts, browser work, and retries all belong to the meter.

Paid plans can continue past included quota with on-demand credits. The site source records that overage is billed at API list prices for the model in use, and that priority or speed configurations can cost more. Purchased credits roll over rather than disappearing at the end of a month. On Teams, the pool is shared, which makes central budgeting possible and also makes one expensive session everybody's problem.

Caching and prompt discipline matter. Cognition's own quota guidance, as summarized in the existing site, recommends precise prompts, removing unnecessary context, avoiding needlessly long sessions, and staying on one frontier model so cache reuse can help. That is not a bag of micro-optimizations. Autonomous loops repeatedly carry context, and a vague task pays for exploration before it pays for implementation.

Teams pricing without the misleading per-seat shortcut

Teams has an $80 monthly minimum. Full seats cost $40 each and include a Pro-equivalent allowance plus Devin Desktop access. Flex seats are free and unlimited, but they do not carry included quota; their work draws from the shared on-demand credit pool. This makes the plan unusually sensitive to how many people drive the product every day.

A ten-person team does not automatically need ten full seats. If two developers use Devin every day and eight only assign occasional tasks, two $40 full seats plus free flex membership can be rational. The occasional work then consumes shared credits. The reverse is also true: declaring everyone flex to avoid seats can be more expensive if recurring usage is paid entirely at on-demand rates.

The right allocation requires a month of person-level usage evidence. Give full seats to predictable daily users, flex seats to occasional requesters, and keep a shared budget for the latter. Revisit the split after a release cycle because autonomous usage is often episodic. A team may look light in an ordinary month and become extremely bursty during a migration.

What happened to Agent Compute Units

Agent Compute Units, usually shortened to ACUs, were Cognition's abstraction for work performed by Devin. Older public pricing attached dollars and plan allowances directly to that unit. The advantage was conceptual: buyers saw an agent-work meter instead of a raw list of model tokens and environment events. The disadvantage was comparability. An ACU did not equal a portable amount of accepted software, so a team still had to learn how its repository converted ACUs into useful changes.

The self-serve model changed in March 2026. Free, Pro, Max, and Teams now present included quota and on-demand credits. Legacy ACU value migrates to credits at the same dollar value, according to the billing coverage in the site. Legacy Core users moved to Free and can spend remaining value; buying more can require moving to the current Teams structure.

Enterprise is the qualification that prevents a clean headline. Cognition's enterprise documentation can bill Agent Compute Units at a volume and rate set in the order form. For an enterprise buyer, the ACU is still a contractual meter. For a new $20 Pro buyer, it is historical vocabulary and not the unit shown on the current public plan ladder.

The real economics: cost per accepted task

Autonomous coding shifts labor rather than removing it. Someone selects and scopes the ticket, supplies repository knowledge and access, reviews the returned change, reruns important evidence, and handles rework. A $20 subscription can be expensive if it creates a queue of plausible pull requests nobody trusts. A $200 plan can be cheap if it reliably completes a repeated migration that previously consumed days of senior engineering time.

accepted-task cost =
  subscription allocation
  + on-demand model usage
  + scoping time
  + review time
  + rework time
  + expected incident cost

throughput = accepted tasks / elapsed calendar time

Count acceptance strictly. A pull request opened is not an accepted task. A branch that passes the agent's tests but needs a human rewrite is not an accepted task. A change merged with an unreviewed deletion is not evidence of cheap work. The task counts when the responsible engineer can explain it, the agreed checks pass in the relevant environment, and the change satisfies the original outcome.

Parallel sessions increase calendar throughput only while review keeps up. If five sessions finish overnight and one reviewer needs two days to inspect them, the product has moved the queue rather than removed it. Put a work-in-progress limit on autonomous tasks just as you would on human work. The correct limit is determined by review capacity, not the maximum number of agents the plan can start.

Which Devin plan is worth paying for

The Free plan is the right answer until the workflow is proven. The $20 Pro plan is cheap enough that one accepted medium task can justify it, but that does not make the product automatically valuable. A disciplined pilot should show which task classes succeed, how many minutes of review each creates, and how often quota stops productive work.

Max should solve a measured capacity problem. If the agent is working on well-scoped tasks and the only recurring failure is the daily boundary, removing that cap has a clear value. If the real failure is vague assignments, missing environment setup, or expensive rework, ten times the monthly price buys more of the failure.

Teams and Enterprise are operating-model purchases. Central billing, shared context, admin analytics, SSO, deployment controls, and support can be worth more than raw agent capacity. Price those requirements explicitly. Comparing an enterprise order form with a $20 individual CLI subscription while ignoring policy and support is not a serious comparison.

Run a pricing pilot that survives procurement

A fair Devin pilot uses repeated real work and a constant acceptance bar. Do not compare one polished demo with the worst incident in your backlog. Do not compare a Devin run with a local agent while treating human steering as free. Record all person-minutes on both sides.

Questions people ask

Keep reading