Devin
Cognition unveils Devin, billed as the first AI software engineer, resolving 13.86% of SWE-bench issues end to end.
An agent with its own shell, editor and browser that plans and executes long coding tasks. Cognition reported 13.86% on a random 25% subset of SWE-bench versus 1.96% previous unassisted best, a launch that set off the 'AI engineer' product wave.
- Date
- Tuesday, 12 March 2024
- Lab
- Cognition
- Kind
- product
- Access
- app only
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench (25% random subset) | 13.86% vs 1.96% prior unassisted SOTA; baselines were given the files to edit | company |
Early access via waitlist at announcement. The headline number is Cognition's own, on a subset, with a different assistance setup than baselines.
Sources
This record was checked against its sources on 6 October 2026. How we check