Devin

Devin is a fully autonomous AI software engineer built by Cognition AI, founded in 2023 and backed by Founders Fund, Elad Gil, and others with over $174M raised. Where tools like Cursor and GitHub Copilot augment human developers in real time, Devin operates independently -- you assign it a task via Slack, GitHub issue, or the Devin web interface, and it plans, codes, tests, debugs, and submits a pull request without supervision. Devin operates in a persistent sandboxed cloud environment with its own terminal, code editor, and browser. It can read documentation, search Stack Overflow, install dependencies, run tests, interpret error logs, and iterate autonomously. The knowledge base feature lets it learn your codebase conventions and architectural patterns, improving output quality over time. Devin 2.0, launched December 2025, dropped entry pricing from $500/month to $20/month and introduced parallel task execution -- spinning up multiple Devin instances simultaneously on a single codebase. Cognition AI acquired Windsurf (formerly Codeium) for approximately $250 million in 2026, putting Devin and Windsurf under the same parent company. One positions autonomous agent execution, the other IDE-first agentic coding -- a strategic consolidation of both ends of the AI coding spectrum. Pricing uses Agent Compute Units (ACUs) -- Cognition''s normalised measure of compute resources (virtual machine time, model inference, networking). One ACU equals approximately 15 minutes of active Devin work. Core plan: $20/month, pay-as-you-go at $2.25/ACU. Team plan: $500/month with 250 ACUs included at $2.00/ACU plus API access. Enterprise: custom pricing with private cloud deployment, SSO, and dedicated SLA. Devin for Terminal launched April 2026 -- a CLI agent that keeps running after you close your laptop, extending Devin''s autonomous execution to terminal-first workflows. Cognition reports Devin successfully completes approximately 75% of assigned tasks, with the remaining 25% requiring human intervention. Best suited to well-scoped, testable tasks: bug fixes with clear reproduction steps, test coverage expansion, library migrations, boilerplate generation, and documentation updates. G2 rating: 4.2/5. Trustpilot (Cognition): 3.0/5 as of March 2026 -- recurring themes include task failures without clear explanation and compute limits at the $20/month tier.