An Open Coding Agent Catches Up to the Paid Ones

OpenHands, the open-source autonomous coding agent formerly known as OpenDevin, has reached version 1.0. That number usually signals a project deciding it is ready to be trusted with production work, and this release is built around exactly that claim. It ships with Docker sandboxing, built-in security policies, resource limits and a plugin system, the kind of plumbing that turns an interesting demo into something you might actually point at a real codebase.

The headline figure is the one that will get attention. OpenHands 1.0 autonomously completes roughly 68 percent of tasks on SWE-bench Verified, the standard test set of real GitHub issues. That puts it in the same conversation as the commercial coding agents that bill by the task. For a project developed in the open on GitHub under an MIT license, closing that gap is the story. For most of the past two years, the open option was the one you settled for.

The cost math is the quiet revolution

Benchmarks make headlines, but the economics may matter more. According to a hands-on writeup of the release, a self-hosted OpenHands stack costs somewhere between 20 cents and about a dollar per resolved task at typical H100 GPU rates, depending on which model you put behind it. Paired with a smaller open model, Devstral 24B, it reached 46.8 percent on SWE-bench Verified, which the writeup notes matches what the funded commercial product Devin 2.0 has publicly reported.

Sit with that comparison for a second. A model you can run on your own hardware, with no data leaving your network, roughly matching a well-financed proprietary agent. The reason self-hosting mattered was never bragging rights. It was that sending your source code and credentials to someone else's servers is a hard sell in a lot of companies. Until now the price of keeping everything in-house was a noticeably weaker agent. That trade looks a lot softer today.

Where it fits

This is part of a broader thread we have followed, where small and open systems keep punching above their weight. A compact research agent recently outperformed Claude and GPT at reproducing published science, and the same pressure is now showing up in everyday software work, the kind that fuels the recurring worry about whether coding jobs are next.

I would not oversell it. A 68 percent score means the agent still gets a third of these curated problems wrong, and real production bugs are messier than a benchmark. The security rails are welcome precisely because an agent with shell access and a plugin system is a liability if it goes sideways. But the direction is unmistakable. When the free, inspectable, self-hostable option performs like the paid one, the question for a lot of teams stops being which vendor to trust and becomes whether they need a vendor at all.