Devin
Marketed as an 'AI software engineer' that plans, writes, tests, and opens PRs for assigned engineering tasks with a cloud sandbox and its own dev environment.
This is where the depth lives — everything below is evidence-graded
How the product is positioned
Marketed as an 'AI software engineer' that plans, writes, tests, and opens PRs for assigned engineering tasks with a cloud sandbox and its own dev environment.
Our summary of the vendor’s positioning — verbatim claim not yet recordedWhat we found
Level 1 · assistive
Seeded at the rubric default. No independent evidence of end-to-end task completion has been reviewed yet, so the lowest grade applies until it is.
Any human checkpoint here is a product setting, not a regulatory requirement — confirm how it is configured for your deployment.
Graded 2026-08-01 · re-check 2026-11-01
Procurement detail
- Deployment
- saasvendor documentation
- Compliance
- SOC 2 Type IIreport dates not disclosed — request before signing
- Integrations
- GitHub claim unverifiedGitLab claim unverifiedSlack claim unverifiedLinear claim unverifiedJira claim unverified
- Pricing
- usageTeam plan ~$500/mo for a fixed compute allotment plus usage (ACUs); Enterprise custom.
- Human checkpoint
- Requiredproduct setting
- Named customers
- Not independently verified beyond vendor-cited logosfrom cited sources
Will they still be here?
- Funding
- Raised $400M+ at $10.2B valuation (Sept 2025), two months after acquiring Windsurf
- Founded
- 2023
- HQ
- San Francisco, USA
- Ownership
- $400M raised following Windsurf acquisition; valued at $10.2B (Sept 2025); acquired Windsurf's IP/product/team (July 2025) after a Google/DeepMind licensing deal took Windsurf's original leadership
Sources
Logo retrieved 2026-08-01from the vendor’s own site · trademark of Cognition
- cognition.comIndependent · retrieved 2026-08-01
- cnbc.comIndependent · retrieved 2026-08-01
Where it’s weak
Self-reported 13.86% SWE-bench resolution at 2024 launch is the last independently comparable number published; Cognition has not released updated standardized benchmark scores since, even as Claude/GPT-5 class models report 70%+ on SWE-bench Verified — the 'autonomous engineer' framing is not supported by current independent evidence and real-world reporting describes it as still requiring significant human review/oversight in production.
Corrections
Vendors may submit factual corrections free of charge. We correct facts and publish the change. We do not accept payment to change a grade. Submit a correction →