Devin

What it is

Devin is Cognition’s AI software engineer, and since June 2026 it is also the name of the editor many teams knew as Windsurf. One subscription now covers three surfaces: Devin Cloud, the autonomous agent you hand a ticket to in Slack, Linear or Jira and that comes back with a pull request; Devin Desktop, the former Windsurf IDE; and the Devin CLI, which starts a task locally and hands it off to the cloud. Devin Review (AI pull-request review) and DeepWiki (auto-generated docs for a repository) sit on the same account.

Cognition bought Windsurf’s product, brand and remaining team in July 2025, after Google hired its founders. On 2 June 2026 the editor shipped as Devin Desktop through an over-the-air update: same account, plan, settings, extensions and keybindings, with no reinstall. The default screen became the Agent Command Center, a board of local and cloud agent sessions with the full IDE one click behind it. Desktop also runs any agent that speaks the open Agent Client Protocol (ACP) next to Devin, so a team can keep Codex or Claude’s agent inside the same window. Windsurf’s old in-editor agent, Cascade, stayed available only until 1 July; scripts and rules that called it by name had to be repointed.

The company behind it is not going away. Cognition raised over $1B at a $26B valuation on 27 May 2026, then over $2B at $48B on 8 September (Andreessen Horowitz and Accel leading), and reported a $1B annualised revenue run rate on 25 September.

Why ops teams run into it

Two ways in. Someone on the GTM-engineering or RevOps-engineering team was a Windsurf user and now finds Devin on their screen and asks whether to stay. Or an engineering lead pitches Devin Cloud as a way for the ops team to file “build me this Salesforce-to-warehouse sync” as a Linear ticket and get a pull request back without a developer.

The second case is the real decision. Devin’s distinct strength is delegation: tag @Devin in Slack or Microsoft Teams, assign a Jira or Linear issue, and it works in its own cloud machine while you do something else. Cursor and GitHub Copilot are built first for the person at the keyboard; Devin is built first for the person who wants to hand work off.

Pricing reality

List prices, checked on devin.ai/pricing and the Devin billing docs today:

• Free — $0, a light quota, limited model choice, unlimited Tab completions.
• Pro — $20/month, Pro-level daily and weekly quota, frontier models, Devin Cloud.
• Max — $200/month, a much larger weekly quota and no daily cap.
• Teams — $80/month minimum, up to 200 members. Full seats are $40/month each and get a Pro-equivalent quota plus Devin Desktop. Flex seats cost nothing, draw only from the team’s shared pool of on-demand credits, and do not include Desktop.
• Enterprise — custom, still billed in Agent Compute Units (ACUs), with SAML/OIDC SSO and VPC deployment.

Three details decide the real number. Cognition does not publish quota sizes — plans say “light”, “increased” and “significantly higher” — and quotas are counted in tokens, so a long session over a large repo spends far more than a short edit. Usage past quota is billed at the model provider’s API list price, and purchased credits never expire. Free models do not count against quota; Cognition’s own SWE-2 is free in Desktop and the CLI until 16 October 2026, after which that line of the bill changes.

A worked band for a 10-person ops group: three builders on full seats ($120) plus seven flex seats for people who delegate the occasional ticket from Slack comes to $120 a month before credits. Ten full seats is $400, the same as Cursor Teams at $40 per user, double GitHub Copilot Business at $19, and well under ten Claude Team premium seats at $100 each. Devin Review is usage-based after a two-week trial and stays free on open-source repositories.

Enterprise buyers can also ask about the AI Productivity Guarantee Cognition announced in June: per press coverage, credits of up to $10M if Cognition’s own estimate of engineering time saved over an annual contract falls short of what you spent. The measurement is Cognition’s, so get the method written into the order form before you count on it.

Best for

An ops engineering lead, GTM engineer or RevOps engineer who wants to hand well-scoped tickets to an agent from Slack, Linear or Jira, with a few builders in the IDE and a wider group delegating on flex seats. It wins on that split: one account covers the people who code and the people who only file the request.

Not for

A team that needs a predictable monthly number and can’t tolerate unpublished quotas. Also not for an ops team without a reviewer: Devin opens pull requests on its own, and somebody who reads code still has to approve them.

Versus the alternatives

• GitHub Copilot — the market-share leader, $19 per Business seat. Pick it when your code already lives on GitHub and most work is completion and occasional agent help; it is cheaper and already under contract in most companies.
• Cursor — the strongest IDE challenger, $40 per user on Teams. Pick it when the work happens at the keyboard and you want the editor that most MCP-heavy builders standardise on.
• Claude Code, included in Claude Team seats ($20 standard, $100 premium, annual) — the fastest-growing tool in the segment. Pick it when the work is long, terminal-first sessions over a whole repo or MCP-connected systems.
• OpenAI Codex, bundled with ChatGPT Business — pick it when the company is standardised on OpenAI.

Pick Devin over all four when hand-off — ticket in, pull request out — matters more than in-editor speed. See best AI coding assistants and alternatives to Cursor.

Watch-outs

• Quotas you can’t see turn into API-priced overage. With no published allowance and overage at list API rates, a week of long cloud sessions can cost more than the seat. Guard: set a default session spending limit under Settings → Usage & limits, keep auto-reload off for the first month, and run a two-week pilot that logs spend per task before buying full seats for everyone.
• The product moves under you. In six months Cognition retired its Core and Team plans (April), renamed the IDE and retired Cascade (June–July), and shipped three model generations (SWE-1.6, 1.7 and 2). Guard: keep workspace rules in the new .devin/ directory, don’t hard-code agent or model names into CI, and recheck pricing after 16 October when the free SWE-2 window closes.
• An autonomous agent with repository and secrets access. Devin Cloud runs unattended with whatever credentials you give it, and a confident pull request can still be wrong. Guard: give it a scoped service account and only the secrets the task needs, keep branch protection with a required human approver, and start it on internal tooling rather than production revenue systems.