Devin
CognitionAutonomous software-engineering agent that runs in a cloud sandbox with its own shell, browser and editor, taking a task from specification through to a pull request. It is reached through a web interface, a CLI, a desktop app and an API, and runs on Cognition's own in-house model.
A previously cited SWE-bench figure is no longer on the page it came from, so the capability band rests on the vendor's own report and a third-party leaderboard row. Verified 2026-08-13 via the Devin documentation and Cognition's SWE-bench technical report.
Openness
1 high confidence- source
- closed
- license
- proprietary(SaaS-only)
- self-host
- no
Proprietary subscription SaaS: no source is published and there is no self-hosting path.
- https://docs.devin.ai recorded 2026-08-13
proprietary SaaS coding agent documented across Cloud, CLI, Desktop, Enterprise, API and Federal; no source distribution and no self-host path
- https://devin.ai/pricing recorded 2026-08-13
paid tiers - Free $0, Pro $20/mo, Max $200/mo, Teams $80/mo plus $40 per dev seat, Enterprise via contact sales; proprietary SaaS throughout
Adoption
3 low confidenceWidely adopted commercial coding agent with paid Individual, Teams and Enterprise tiers, but Cognition discloses no hard user count. The level rests on reported commercial traction rather than a counted download or user figure.
- https://devin.ai/pricing recorded 2026-08-13
individual, team and enterprise commercial tiers indicating broad paid adoption; no user or seat count published
Capability
3 medium confidenceDevin 2.0 scores 45.8% SWE-bench Verified and 35.6% Full on the coding-agent leaderboard, as a single agent with no human in the loop, and the figure is independently corroborated. That is mid-tier, well below the OpenHands anchor at the top of the field. Cognition's own earlier technical report gave 13.86% on a 570-problem subset. The Tembo post cited beside the leaderboard no longer carries the number, so the leaderboard is what the score rests on.
- https://cognition.ai/blog/swe-bench-technical-report recorded 2026-08-13
Cognition team, 2024-03-15 - "In SWE-bench, Devin successfully resolves 13.86% of issues, far exceeding the previous highest unassisted baseline of 1.96%"
- https://www.tembo.io/blog/devin-alternatives-2025 recorded 2026-08-13
Previously cited for Devin 2.0 at 45.8% SWE-Bench Verified; the figure no longer appears anywhere on the page.
- https://awesomeagents.ai/leaderboards/swe-bench-coding-agent-leaderboard/ recorded 2026-08-13
row 7 - Devin 2.0, proprietary, SWE-Bench Verified 45.8%, SWE-Bench Full 35.6%, Cognition AI, unassisted standard eval
Verified 2026-08-13