Devin vs OpenHands: Managed Autonomy vs Self-Hosting
Devin 2.0's ACU pricing vs OpenHands' self-hosted open-source model — a clear-eyed look at cost, control, and benchmark caveats.
The “AI software engineer” category has shaken out into two distinct philosophies: pay a vendor to manage everything, or run the stack yourself and pay only for model tokens. Devin from Cognition and OpenHands from All Hands AI are the clearest poles of that divide. Neither is obviously right. What matters is which tradeoffs map to your team’s constraints.
What Each Product Actually Is
Cognition’s Devin is positioned as the world’s first AI software engineer — it independently plans, executes, and iterates on complex engineering tasks requiring thousands of decisions.
It operates in its own sandboxed environment with a complete setup — shell, code editor, and browser — just like any human developer would use. The key word is managed: Cognition controls the infrastructure, the runtime, and currently the underlying model stack.
OpenHands is currently the closest open-source equivalent to a “final form” AI software engineer — the community’s answer to proprietary tools like Devin.
Within a comprehensive GUI environment that integrates a code editor, terminal, browser, and live preview, OpenHands’ agents can autonomously plan tasks, execute commands, write code, and verify results — a closed loop from requirements to deployment.
Released under the permissive MIT license, it is a community project with more than 2.1K contributions from over 188 contributors.
Pricing: A Tale of Two Models
Devin’s pricing restructured significantly in April 2025. Devin now offers a $20/month Core plan, a $500/month Team plan, and a custom-priced Enterprise plan.
It charges based on consumption of ACUs — Agentic Computing Units — which are Cognition’s normalized measure of the resources Devin uses while actively working: virtual machine time, model inference, and networking bandwidth.
For the starting price of $20, customers gain access to 9 ACUs.
The pay-as-you-go model offers additional ACUs at $2.25 each.
The cost trap with the ACU model is opacity. ACUs reflect the complexity and duration of tasks like planning, debugging, context gathering, code execution, and browser actions — and tasks do not consume ACUs equally; longer workflows use more ACUs, while idle time uses none. Teams with heavy, multi-hour task loads can find themselves well above the sticker price quickly. At the Team tier, the $500/month plan targets medium-sized companies needing an AI coding assistant for multiple projects, providing 250 ACUs, with additional ACUs available for $2 each.
OpenHands’ cost profile is structurally different. OpenHands uses LiteLLM as its model abstraction layer, so it works with any provider LiteLLM supports: Anthropic Claude, OpenAI, Azure OpenAI, Gemini, and any OpenAI-compatible self-hosted endpoint. You pay for model tokens, not a proprietary compute unit — which means costs are directly tied to actual LLM consumption, not to Cognition’s infrastructure margin. Self-hosted OpenHands runs entirely in your infrastructure with no external data transmission.
OpenHands Cloud also comes with source-available features for teams that want the managed path. For enterprise-scale deployments with strict data requirements, large enterprises can work with All Hands AI to self-host OpenHands Cloud in their own VPC via Kubernetes.
Control and Customization
This is where the gap is most pronounced. Devin is a closed system. You interact via its interface or Slack; you don’t control the runtime, the model, or the scaffolding. That’s fine if you want zero operational overhead, but it means you cannot:
- Swap in a cheaper or more capable model when one ships
- Run tasks on-premises for compliance reasons
- Instrument the agent loop for audit or debugging
- Integrate deeply with internal tooling that Cognition hasn’t built for
OpenHands inverts every one of those constraints. OpenHands runs in your environment — whether on-prem or private cloud — so your code never leaves your control, and since agents operate inside secure, isolated execution environments with full auditability, it is built for organizations with strict data security, compliance, and governance requirements. Teams who want to pair it with Claude Code or Cursor workflows have clean integration paths at the model layer. The tradeoff is that you own the ops burden: provisioning containers, managing sandbox lifecycle, keeping pace with upstream releases, and debugging when the agent loop fails in unexpected ways.
Benchmark Performance — Read With Caution
SWE-bench scores are the standard currency for comparing these tools, and both have strong numbers — but the numbers require careful handling.
Devin’s headline benchmark figure dates to its early release. On SWE-bench, Devin resolves 13.86% of real GitHub issues end-to-end — a 7x improvement over previous AI models — though independent testing shows 15-30% success rates in practice.
OpenHands has posted considerably higher numbers, particularly as the underlying models have improved. On SWE-Bench Verified, the OpenHands SDK achieves a 72% resolution rate using Claude Sonnet 4.5 with extended thinking.
OpenHands is also a top-performing agent on LiveSWEBench — a contamination-free benchmark — and ranks first on Multi-SWE-Bench, which evaluates across 8 different programming languages.
However, headline SWE-bench scores do not transfer cleanly to real-world tasks. The same methods achieve substantially higher scores on SWE-bench compared to their performance on SWE-bench-Live, despite both benchmarks targeting the same issue-resolving task — with recent state-of-the-art agents reporting resolved rates exceeding 60% on SWE-bench Verified, while the highest resolved rate on SWE-bench-Live is only 19.25%.
This highlights the challenges of constructing a benchmark that objectively measures an AI system’s ability to resolve arbitrary and previously unseen issues, and raises concerns about potential overfitting to SWE-bench.
Both Devin and OpenHands are subject to this caveat. Treat leaderboard numbers as a proxy for potential, not a prediction of production success rate on your codebase.
For further context on how autonomous coding agents compare across the board, see our Devin vs Claude Code comparison and the Cursor vs GitHub Copilot breakdown for the assisted-coding side of the spectrum.
Who Each Tool Is Actually For
Devin makes sense if your team values zero ops overhead above all, you’re operating at the Team tier or above with a steady task volume, and your compliance posture allows third-party infrastructure to touch your codebase. The ACU model becomes legible once you’ve run a few weeks of real tasks and know your consumption patterns.
OpenHands makes sense if data sovereignty is non-negotiable, you want model flexibility as the frontier keeps moving, or you’re technically capable enough to own the self-hosted stack. It is used by platform teams automating engineering workflows at scale, teams building internal developer platforms or AI-powered tooling, and enterprises managing large or legacy codebases and mission-critical software systems.
Teams evaluating either should also look at Factory and GitHub Copilot before committing — the autonomous engineering space is moving fast enough that your shortlist from three months ago is already stale.
Bottom Line
Devin buys you a fully managed experience at a predictable (if opaque) ACU cost, backed by a commercial team that owns the reliability contract. OpenHands gives you full control, model agnosticism, and MIT-licensed auditability at the cost of real operational overhead. Neither benchmarks nor marketing copy will tell you which one survives contact with your actual backlog — run both against a representative slice of your real tickets before committing budget.