Devin is the agent people mean when they say “AI software engineer.” It is made by a company called Cognition, and it is a different category from the consumer agents Meta and OpenAI shipped in 2026. Dots and Muse shop and send email. Devin writes, tests and ships code. If you are not a developer, most of what follows will not apply to you, and that is the point.
Devin runs in a sandbox virtual machine with a browser and a terminal. It integrates with GitHub, GitLab, Linear, Jira, Slack, Notion and Sentry. You give it a task, it explores the codebase, it makes the change, it runs the tests and it opens a pull request for you to review. The pitch is an engineer that does not get tired, not a chatbot that explains how.
How Cognition prices it now
Devin launched at $500 a seat with no free tier. That model is gone. Cognition restructured self-serve pricing into Free, Pro, Max, Teams and Enterprise tiers. Pro runs $20 per seat a month. Max runs $200.
Teams starts at $80 a month plus $40 per full developer seat. Enterprise is custom.
The $20 Pro entry is a door fee, not the real cost. Devin is metered by ACUs, and heavy autonomous runs bill on top of the seat price.
ACU stands for Agent Compute Unit. One ACU is roughly fifteen minutes of active autonomous work, so an hour of Devin runs about $9 on the Core plan at $2.25 per ACU. The Team plan at $500 a month includes 250 ACUs, which works out to around 62 hours of engineering time before you pay extra. Seats are unlimited across plans. You pay for the work, not for the headcount.
The older $500 entry that scared people away is gone, but the real bill still scales with how much Devin runs. A tightly scoped task can finish for a few dollars. A vague one can burn far more. The meter is the product.
Where it earns its keep
Devin performs best on well-scoped tasks with a clear way to verify the result. A tight spec. A pointer to the exact files. A test suite or a CI check that proves the change works. Hand it those and it implements instead of guessing.
Cognition’s own growth story backs this up. Its annual recurring revenue climbed from about $37 million in May 2025 to $492 million in May 2026, and reached nearly $900 million by August 2026. Mercedes-Benz compressed an eight-month legacy modernization project into eight days using Devin. The pattern repeats: a queue of defined work, not an open-ended wish.
Devin also learns. After a complicated task, it can write a “skill file,” a saved set of instructions it follows to do the same job faster next time. The longer you use it, the better it gets at your specific workflow. That is the kind of compounding a human team also builds, just stored as text.
The failure mode is the opposite. A vague, open-ended prompt lets Devin explore, backtrack and re-run tests, burning ACUs the whole time. A single feature can cost several dollars if you let it wander. Scope tightly and the same feature finishes in fewer ACUs, which on Core directly lowers the bill.
What it is not
Devin is not a personal assistant for everyday life. It will not book your flight or cancel your subscription. It is an engineering tool, and the skills that make it useful are software skills. If you were hoping for the agent from the consumer demos, this is the wrong product and the wrong price.
It is also not a human-in-the-loop editor like Cursor or GitHub Copilot. Those sit beside you and wait for your keystrokes. Devin runs on its own and comes back with a result. That autonomy is the feature and the risk. You trade oversight for throughput, which only pays off when the task is boring enough to delegate and important enough to verify.
Because ACUs track runtime, the lever is scoping, not seat count. Leave Devin a finished UI design and it spends compute implementing the screen instead of inventing the layout. Give it a vague goal and a large codebase and it will spend compute finding the door.
How Devin compares to the others
So Devin is the odd one out. Dots and Muse aim at everyday tasks. Devin aims at code. But that focus is its strength. And it is why the price model looks nothing like the others.
Devin wins on a clear job. It ships code you can verify. But it loses on audience, since you must write software to use it. And it loses on predictability, since the meter moves with every run.
Who should try it
The Free tier makes the trial cost nothing, which is the right entry for an individual developer with a real backlog. Pro at $20 plus on-demand ACUs suits anyone running scoped autonomous tasks who wants to see the meter before committing. Team at $500 a month makes sense once steady use approaches the 250 included ACUs, because the per-ACU rate drops and billing gets predictable.
Skip Devin if you are not writing or shipping code. Skip it if your tasks are too fuzzy to verify, because an agent that cannot be checked is an agent you will not trust with production. And skip it if a $20 editor already covers your workflow, since those tools cost less and keep you in the loop.
My read is that Devin is the most honest of the 2026 agents about what it is. It is a metered engineer, not a friend. The $500 entry is gone, but the real bill still scales with how much it runs. Treat it like contract help: cheap to try, worth it only when the work is clear.
