OpenHands 1.0: a self-hosted coding agent hits 72% on SWE-bench
The real 1.0 is a safety story: a Docker sandbox that pauses risky actions, on your own models at ~5% of Devin's cost. Plus new GPT-Image API models.
Copy markdown
72% on SWE-bench, and you host it
OpenHands 1.0 solves 72% of SWE-bench Verified paired with Claude Sonnet 4.5, and 68% with open-weight Qwen3-Coder-480B — against Devin's self-reported 77.8%. That's near-frontier autonomy running on infrastructure you control instead of a vendor's cloud.
The real 1.0 is the sandbox
Every action runs in a Docker sandbox with per-container CPU and memory caps and non-root execution, while an LLM analyzer scores each step LOW, MEDIUM, or HIGH. High-risk moves — destructive bash, anything touching credentials — pause and wait for your approval. Well-timed, after a week of agent-security incidents.
Bring your own model, ~5% of Devin's cost
It's MIT-licensed and wired through LiteLLM, so it drives 100+ backends — Claude, Gemini, or fully local via Ollama with Qwen3-Coder or DeepSeek V3.2 — at roughly 5% of Devin's per-task cost. Build this weekend: a coding agent that never sends your code off your own machine.
Elsewhere: ChatGPT Images 2.5 adds two API models
OpenAI shipped Images 2.5 with two new API models, GPT-Image-2.5 Flare and Sunburst, plus faster generation, steadier multi-edit consistency, and an in-canvas Sketch tool for drawing prompts directly.