Devin

Devin is fundamentally different from code assistants like GitHub Copilot or Cursor. Those tools work as inline suggestions within your editor, requiring you to remain in the driver's seat at all times. Devin operates as a fully autonomous agent with its own development environment — you assign it a task, and it independently plans, executes, tests, and delivers the result. Think of Copilot as a smart autocomplete, while Devin acts more like a junior engineer you can delegate entire tasks to.

Devin can clone any Git repository and autonomously explore the codebase to understand its architecture, dependencies, and conventions. It reads READMEs, examines file structures, traces function calls, and reviews existing tests before making any changes. This allows it to implement new features that naturally fit the existing code style and project patterns rather than producing generic, out-of-context code.

Devin is designed for production workflows with human oversight. It generates pull requests rather than directly committing to main branches, enabling standard code review by your engineering team. For critical systems, teams typically use Devin for well-defined tasks like test writing, dependency upgrades, or documentation, and treat its output as a starting point for human review. Many companies use it successfully for internal tooling, automation scripts, and greenfield projects.

ACUs no longer appear on Devin's self-serve pricing. Checked on 30 October 2026, the page does not mention ACUs or credits at all; it describes each plan by quota instead. Free is $0 with a light quota and limited model availability, Pro is $20 a month with increased quotas, Max is $200 a month with significantly higher quotas, and Teams is $80 a month plus $40 for each full developer seat for up to 200 users. Enterprise is quoted on request, so a contract there can still set its own rate. Paid-plan usage refreshes daily and weekly, though no numeric quota is published for any plan, and there is no annual billing option. The $500 a month team plan and the 250-ACU Core plan that older comparison articles still quote have both been retired.

Devin excels at well-defined engineering tasks with clear success criteria: writing automated tests for existing code, migrating from one library or framework to another, implementing CRUD features based on API specifications, setting up CI/CD pipelines, fixing known bugs with reproduction steps, refactoring legacy code modules, and writing technical documentation from source code. Tasks that require deep product intuition or undefined creative direction are better suited for human engineers.

Cognition announced Devin on 12 March 2024 under the headline "the first AI software engineer." That launch post reported Devin resolving 13.86% of SWE-bench issues end to end against a previous state of the art of 1.96%, and noted that earlier models managed only 4.80% even when handed the exact files to edit. Those are the figures from the announcement rather than current benchmark standings. Cognition's site now lives at cognition.com, and the older cognition.ai address redirects there.

Yes. Cognition released Devin CLI on 27 April 2026 as a local coding agent that runs in your shell with access to your codebase, your tools and your environment, installed with the one-line script at cli.devin.ai/install.sh. When a task outgrows your machine you can hand the session to a cloud agent that has its own computer, so the work continues after you close your laptop. The CLI lets you choose which model to run, and Cognition names Opus 4.7, GPT-5.5 and its own SWE-1.6 among the options.

Devin CLI reads .devin/config.json, which you commit so the whole team shares it; at project level it accepts permissions, read_config_from and hooks. Your personal overrides go in .devin/config.local.json, which is gitignored. MCP servers are defined in .devin/mcp_config.json, with .devin/mcp_config.local.json for local ones, and project skills live under .devin/skills with a SKILL.md inside each folder. The read_config_from setting can also import rules from files other tools already left in the repo, such as AGENTS.md. Model choice is a user-level setting, not a project one.

Teams starts at $80 a month for the plan itself and adds $40 for each full developer seat, so three developers come to $200 a month and five come to $280, up to a ceiling of 200 users. Below that, Pro at $20 a month covers one member, and Max at $200 a month covers one member with significantly higher quotas — so a two-person team on individual Pro seats pays $40 rather than $160. Prices checked 30 October 2026. What the plans do not publish is how much work each quota actually buys, so budget from a trial rather than from the sticker price.