Last updated: 2026-06-12
Devin by Cognition is an autonomous AI software engineer that plans, codes, tests, and ships pull requests from $20/month plus $2.25 per Agent Compute Unit.
Devin, Cognition's autonomous AI software engineer, costs $20 per month for the Core plan plus $2.25 per Agent Compute Unit, or $500 per month for the Team plan with 250 ACUs included. It runs in a sandboxed VM with its own shell, editor, and browser, autonomously planning, coding, testing, and opening pull requests for bounded engineering tasks such as legacy code migrations.
Devin is an autonomous AI software engineer built by Cognition, a company founded in August 2023 by Scott Wu, Steven Hao, and Walden Yan. Launched in March 2024, Devin was marketed as the first AI agent that works as a software engineer rather than as a code-completion tool, resolving 13.86% of real GitHub issues end to end at launch, far above the previous best of 1.96%. It runs inside its own sandboxed virtual machine with a shell, code editor, and browser, so it never executes code directly on a user's computer. Devin operates as a compound system rather than a single model: a planner breaks down the task, a coder writes the implementation, and a critic reviews the result for bugs and security issues before it is shipped. The 2026 version runs on Claude Sonnet 4.5, which Cognition reports brought an 18% improvement in planning and a 12% gain in end-to-end evaluation scores, and added dynamic re-planning so Devin can change strategy mid-task if it hits a roadblock. It also accepts multi-modal input, including UI mockups, Figma files, and screen-recording video, to reproduce and fix visual bugs. Devin's standout use case is legacy code migration: it can ingest codebases written in COBOL, Fortran, or Objective-C and refactor them into Rust, Go, or Python while preserving business logic. It is also used for bounded engineering work such as dependency upgrades, lint fixes, test backfills, and bug fixes, where it opens a pull request, responds to review comments, and can record a QA walkthrough of its own testing. Nubank reported a roughly 12x improvement in engineering hours on migration work after adopting Devin. Pricing has three tiers. The Core plan is $20/month plus $2.25 per Agent Compute Unit (ACU), where one ACU is roughly 15 minutes of active autonomous work covering VM time, inference, and networking. The Team plan is $500/month and includes 250 ACUs at $2.00 each. Enterprise pricing is custom and adds VPC-isolated deployment, SSO, and SOC 2 Type II controls. Devin is delivered through Slack, Microsoft Teams, a web interface, and a VS Code extension. Cognition has grown quickly around Devin: it acquired the Windsurf AI coding IDE in July 2025, raised $400 million at a $10.2 billion valuation in September 2025, and was in talks for a $25 billion valuation by April 2026. Customers reported by 2026 include Dell Technologies and Cisco Systems, alongside SOC 2 Type II certification held since September 2024.
Core plan $20/month plus $2.25 per Agent Compute Unit (ACU, roughly 15 minutes of autonomous work) for usage. Team plan $500/month includes 250 ACUs at $2.00 each. Enterprise plan is custom-quoted and adds VPC-isolated deployment, SSO, and dedicated support. No permanent free tier.
Devin is an autonomous AI software engineer built by Cognition, a company founded in August 2023 by Scott Wu, Steven Hao, and Walden Yan. Launched in March 2024, Devin was the first AI agent marketed as a full software engineer rather than a coding assistant. It works inside a sandboxed virtual machine equipped with its own shell, code editor, and browser, so it never runs code directly on a user's computer. Devin can plan a task, write the code, run tests, debug failures, and open a pull request without step-by-step supervision. In 2026, Cognition added dynamic re-planning, so if Devin hits a roadblock it can change its strategy mid-task instead of stalling. By 2026, Devin's customers included Dell Technologies and Cisco Systems, and Nubank reported a roughly 12x improvement in engineering hours on migration work.
Devin has three plans: Core, Team, and Enterprise. The Core plan costs $20/month and includes platform access, with actual work billed separately at $2.25 per Agent Compute Unit (ACU). The Team plan costs $500/month and includes 250 ACUs at the lower rate of $2.00 each, suited to teams with a steady backlog of tasks. Enterprise pricing is custom and adds VPC deployment, SSO, and dedicated support for compliance-sensitive organizations. One ACU represents roughly 15 minutes of active autonomous work, covering VM time, model inference, and networking. Because billing is usage-based, real monthly costs commonly land between $300 and $500 even for teams on the $20 Core plan. The 2026 pricing reset cut the old $500/month entry price by 96%, making Devin accessible to individual developers for the first time.
Devin's core feature is end-to-end autonomous task execution: it plans, writes, tests, debugs, and opens pull requests for scoped engineering work. It runs in a sandboxed VM with a shell, code editor, and browser, isolating all execution from the user's machine. A standout 2026 capability is legacy code migration, where Devin ingests codebases in COBOL, Fortran, or Objective-C and refactors them into Rust, Go, or Python while preserving business logic. Devin also accepts multi-modal input, including UI mockups, Figma files, and screen-recording videos, to reproduce and fix visual bugs. Before requesting review, it can run the app, click through the affected flow, and send back a recorded QA walkthrough. Since February 2026, Devin supports parallel sessions, letting it work on multiple tasks at once with improved context retention on long-running work.
Devin does not have a permanent free tier; the cheapest plan, Core, costs $20/month. All Devin plans bill actual work separately in Agent Compute Units, so even the $20/month plan requires additional usage spend to complete a task. There is no published trial credit amount on Cognition's current pricing page as of 2026. The Team plan at $500/month bundles 250 ACUs, which can offset the per-unit cost for teams with consistent workloads. Enterprise customers negotiate custom contracts that may include committed-use discounts. Anyone evaluating Devin should budget for ACU costs on top of the subscription fee, since reviewers report typical real-world spend of $300 to $500 a month.
Claude Code is the strongest alternative for teams that want deep repo-wide reasoning, multi-file refactoring, and terminal-based agentic coding at flat subscription pricing rather than per-task ACU billing. Cursor is the best choice for developers who want an interactive AI-native IDE for daily pair-programming rather than asynchronous task delegation. GitHub Copilot remains the safest rollout for large organizations already standardized on GitHub, with three distinct agent surfaces as of 2026. Windsurf, which Cognition itself acquired in July 2025, offers an AI-native IDE experience that sits between Cursor's interactivity and Devin's autonomy. Choose Devin specifically when you want a task handed off completely and returned as a reviewable pull request, and choose the alternatives above when you want to stay in the loop during coding.
Devin is best for engineering managers and platform teams with a backlog of well-scoped, bounded tasks such as lint fixes, dependency upgrades, test backfills, and bug fixes with clear acceptance criteria. It is particularly strong for legacy code migration projects, where companies like Nubank have used it to cut migration engineering hours by roughly 12x. Enterprises that need SOC 2 Type II controls, VPC-isolated deployments, and SSO for autonomous coding agents are a good fit for the Enterprise plan. Devin is not well suited to solo developers on tight budgets, since ACU billing can push real costs to $300-500/month even on the $20 entry plan. It also struggles with ambiguous product requirements and complex architectural decisions, where it can fall into unproductive debugging loops. The right mental model is a capable junior engineer who needs a clear brief and PR review, not a senior architect working independently.
Yes, Devin exposes an API with keys prefixed apk_ that lets external systems delegate tasks to Devin and monitor execution, and this powers community projects such as the devin-mcp server. Devin also supports the Model Context Protocol (MCP) through a built-in MCP Marketplace in Settings, giving it native access to tools like Linear, Datadog, Slack, Figma, Stripe, Zapier, and Airtable. If a needed integration isn't in the marketplace, users can add any custom MCP server over STDIO, SSE, or HTTP transports. By Q2 2026, the broader MCP ecosystem included community servers for GitHub, PostgreSQL, Docker, and Kubernetes that Devin can connect to. Task delegation itself is primarily handled through Slack or Microsoft Teams, where assigning Devin a task in a channel spins up a new session. The API and MCP support make Devin scriptable as part of a larger automated engineering pipeline rather than only a chat-based tool.
Claude Code is generally rated the strongest overall AI coding tool in 2026 for repo-wide reasoning, multi-file refactoring, and terminal-based agentic work, and it runs on a flat subscription rather than per-task billing. Devin, by contrast, is built for asynchronous delegation: you hand it a scoped task via Slack and it returns a pull request, working in its own sandboxed VM with a shell, editor, and browser. On cost, Claude Code's pricing is more predictable, while Devin's Agent Compute Unit billing at $2.00-2.25 per ACU can push monthly spend to $300-500 even on the $20 Core plan. On capability, Devin 2.0 runs on Claude Sonnet 4.5 with an 18% planning improvement and a 12% end-to-end eval gain after that upgrade, but its only published SWE-bench figure (13.86%) dates to its 2024 launch and is far below 2026 frontier scores. Choose Claude Code if you want an interactive agent for complex same-session engineering work, and choose Devin if you want tasks fully delegated and returned as reviewable PRs without your involvement during execution.