By OpenHands 1.0 shipped this week with the piece autonomous coding agents have been missing: a production-grade security sandbox. The open-source agent can now autonomously complete 68% of SWE-bench Verified tasks, competes on cost-adjusted performance with commercial rivals like Devin, and — critically — keeps your code and API keys inside your own infrastructure. If you have been holding off on self-hosting an autonomous agent because the security story was not there, 1.0 removes that excuse.
What 1.0 Actually Ships
The headline features are not new capabilities — they are the hardening that makes existing capabilities production-safe. OpenHands 1.0 introduces a Docker sandbox with configurable resource limits (CPU and memory caps per container), non-root execution via SANDBOX_USER_ID=1000, and a built-in LLM-based security analyzer that rates every action LOW, MEDIUM, or HIGH before execution.
The security analyzer matters most. High-risk actions — destructive bash commands, anything touching credentials — pause the agent and wait for an explicit human signal before proceeding. Previously, running OpenHands in production meant rolling your own safety wrapper or accepting best-effort isolation. Now the framework ships it by default. The 1.0 release also adds a plugin system for extending the agent runtime, and moves CLI tooling to a separate repository for cleaner separation of concerns.
Where 68% Sits on the Leaderboard
The benchmark number requires context. OpenHands paired with Qwen3-Coder-480B reaches 68.0% on SWE-bench Verified at 100 turns. Pair it with Claude Sonnet 4.5 and extended thinking and that climbs to 72%. The top of the 2026 leaderboard hovers around 75–80%, with Devin’s SWE-1.7 model claiming 77.8% on Cognition’s own reported testing.
There is a caveat worth internalizing: the harness alone can swing a score by 15 to 20 percentage points, and vendor self-reported numbers carry a real contamination risk. The more honest comparison is cost-adjusted throughput. OpenHands running on open-weight models like Qwen3-Coder or DeepSeek V3.2 completes similar tasks at roughly 5% of Devin’s per-task cost. That math changes the calculus for teams doing volume work: dependency upgrades, boilerplate generation, test coverage, internal tooling.
| Agent | SWE-bench Verified | License | Est. Cost / Task |
|---|---|---|---|
| OpenHands + Qwen3-Coder-480B | 68% | MIT | ~$0.30 in tokens |
| OpenHands + Claude Sonnet 4.5 | 72% | MIT (harness) | ~$1–3 in tokens |
| Devin (Max plan) | 77.8%* | Proprietary | $200/mo + ACUs |
Running It: 100+ LLM Providers, Including Local
OpenHands uses LiteLLM under the hood, which means switching providers requires no code changes. You can point it at Claude, GPT-6 Astra, Gemini, or a local model running via Ollama. The Ollama integration is particularly useful for teams with sensitive codebases: Ollama exposes an OpenAI-compatible endpoint at http://localhost:11434/v1, so your code, context, and API keys never leave your machine or VPC.
The Docker quick-start in 1.0 looks like this:
docker run -it --rm -e SANDBOX_USER_ID=$(id -u) -e LLM_MODEL=claude-sonnet-4-5 -e LLM_API_KEY=$ANTHROPIC_API_KEY -v /var/run/docker.sock:/var/run/docker.sock -v ~/.openhands-state:/.openhands-state -p 3000:3000 ghcr.io/all-handshq/openhands:1.0Verify the flags against the official OpenHands documentation before deploying — the 1.0 release notes have the canonical configuration. For production, add network binding restrictions and review the hardened Docker Compose guide in the repository.
This Is Not a Side Project
OpenHands has 85,000+ GitHub stars, over 9 million downloads, and 500+ contributors including engineers at AMD, Apple, Google, Amazon, Netflix, NVIDIA, and Mastercard. All Hands AI closed an $18.8M Series A in November 2025, led by Madrona Venture Group with participation from Menlo Ventures, Pillar VC, and Obvious Ventures. An enterprise tier with Kubernetes VPC deployment exists for teams that need managed self-hosting. The MIT license covers the core framework and Docker images.
The practical read: this is the open-source coding agent that has the institutional backing to still be here in two years. That matters for teams making architectural decisions about internal tooling.
The Real Argument for Self-Hosting in 2026
Devin’s current pricing runs Free / Pro at $20/month / Max at $200/month, with ACU credits layered on top — one ACU represents approximately 15 minutes of active autonomous work. For teams running dozens of long-horizon tasks per week, those credits accumulate fast. A McKinsey State of AI 2026 survey found that 32% of organizations had already skipped purchasing at least one software product because they could build the equivalent internally using agentic coding tools.
OpenHands is not a universal Devin replacement. For teams that want a managed, no-infrastructure-overhead service, the commercial options still win on setup simplicity. But for teams that are security-conscious, cost-sensitive, or want to run against local models, OpenHands 1.0 is the first version where the answer to “is it production-ready?” is yes without conditions. The security excuse is gone. What you do with that is on you.
