Devin
Cognition's Devin is a hosted autonomous SWE agent. Each session runs in its own persistent VM with a workspace, browser, and shell, plus Slack and IDE integrations so the agent can be assigned tasks like a human engineer. Cognition also publishes DeepWiki, a separate retrieval system over indexed repos that Devin uses to ground long-horizon work.
Leaderboard Placements
| Benchmark | Best base model | Score | Rank |
|---|---|---|---|
| SWE-bench Verified | Proprietary (Sonnet 4.6 + planner) | 61.7 | #12 / 15 |
| Terminal-Bench | - | - | - |
| Aider Polyglot | - | - | - |
| SWE-Lancer | Proprietary (Sonnet 4.6 + planner) | 32.5 | #4 / 5 |
Distribution
Hosted SaaS. Web app, Slack, GitHub, and Linear integrations. No self-host option.
Model Story
Proprietary model mix. Cognition does not disclose which model serves which step but has stated Sonnet 4.6 is in the rotation alongside an in-house planner.
Pricing
Per-seat subscription with usage limits; team and enterprise plans available.
Who It's For
Teams that want an agent assignable through ticketing systems and willing to trade open-source transparency for managed infrastructure.
Notable Features
- Persistent VM workspaces per task
- DeepWiki repo retrieval system
- Slack and Linear assignment surfaces
- Async task queues
- Code review and PR-author workflow