Devin

Devin is for the tickets you would give to a contractor: a cloud workspace, a browser, a shell, and a pull request at the end, with you reading the PR rather than the process. The daily trait is delegation, and delegation only works with a well-specified task; a vague ticket comes back as a confident PR that solves something adjacent, and you pay for the round trip.

Pick it if you can write tickets a stranger could execute and you have a backlog of them. Pick Claude Code or OpenHands if you want to sit in the loop and correct as it goes.

The risk is the refill. The billing docs describe on-demand credits that top up automatically, which is exactly the wrong default for an agent whose purpose is to work while you are not watching: a loop is a bill, and a loop at 3 a.m. is a bill nobody reads until the invoice. Autonomy plus auto-refill is a meter with no floor and no witness.

Set a hard cap before the first task, and make the cap the admin's job, not the user's. What it does right: a sandboxed cloud VM and a pull request as the only output, so nothing lands without review.

Devin's documented architecture is the right one for unattended work: a hosted VM with its own browser and shell, so failures are contained and the output is a pull request a human reads. The evidence is thin. The only benchmark is 13.86 percent on SWE-bench, self-reported in March 2024, on a 25 percent random subset of the test set, not comparable to public entries. The backbone is undisclosed, so no result can be attributed to scaffold versus model.

The consequence is that every capability claim since 2024 is a claim. What would change the assessment: one current number on a public harness, with the model named.

Cognition behaves like a company. It bought Windsurf in July 2025 and folded it into the brand as Devin Desktop, which bought an editor and a user base in one transaction. In April 2026 it replaced the ACU-based plans with Pro at $20 and Max at $200; repricing that sweeping is a tell that the old unit confused buyers or hid margin, and either way the new sheet is designed to be compared with the labs' own plans.

Acquirer: a hyperscaler that wants an autonomous agent with an editor attached. Position: long the company, short the price sheet, and re-read the plans every quarter.

The demo is a Slack message returning as a pull request. Teams is $80 a month plus $40 per full seat, with on-demand credits in a single pool one engineer can drain before the others log in. Sixty full seats is $28,800 a year plus the pool, and the pool is the number. Free flex seats are the part finance will like, since half of sixty engineers would delegate once a month. Enterprise is by quote, which is where SSO and retention will live.

Onboarding is a repository connection. Approved with conditions: Enterprise terms in writing, a spend cap on the pool, and flex seats for light users.

Devin is the closed box the category was named after. No source, undisclosed models, bring_your_own_model false, local_models false, everything in their cloud, and a CLI that is a remote control rather than a runtime. What I can touch: Devin is an MCP server at mcp.devin.ai, so I can call it from an agent I do own, hand it a task, and get a PR back without opening their web app.

That is an integration surface, not ownership, and it goes dark the day the company does. Grudging respect for the API-first design; it is the only part written for someone like me.