OpenHands 1.0 Ships on 8 September 2026
On 8 September 2026, All Hands AI shipped OpenHands 1.0 — the production-ready release of the autonomous coding agent formerly known as OpenDevin. The headline benchmark result: paired with Qwen3-Coder-480B at up to 100 turns, OpenHands resolves 68 percent of tasks on SWE-bench Verified, the most widely cited evaluation for autonomous software engineering. Paired instead with Claude Sonnet 4.5 and extended thinking, the score climbs to 72 percent. Both results place the open-source agent at or above commercial rivals on the same leaderboard, at a fraction of the infrastructure cost.
What All Hands AI Built
All Hands AI is a San Francisco-based company founded in 2024 that maintains OpenHands as an open-source project under a permissive licence. The project began as OpenDevin, a community experiment in applying language model reasoning to real software engineering tasks — filing and resolving GitHub issues — rather than completing isolated code completion exercises. Version 1.0 marks the transition from a research prototype to a system that engineering teams can deploy in production. The platform is available under an open-source licence with no gate on the underlying agent framework, and the company has not disclosed external funding publicly.
Three Architecture Pillars of the 1.0 Release
Production-Ready Docker Sandboxing
Every task OpenHands executes runs inside an isolated Docker container. The sandbox enforces built-in security policies at the container level, with configurable resource limits on CPU, memory, and network access per agent session. This matters because autonomous coding agents frequently need to execute the code they write — running tests, installing packages, starting services, and verifying output — and doing that on a shared host without isolation creates serious security and stability risks. OpenHands 1.0 makes isolation a first-class system guarantee rather than a documentation recommendation.
Bring Your Own Model
OpenHands 1.0 ships with a BYOM framework that lets engineering teams connect any language model API or locally hosted model to the agent harness. Model selection is a configuration decision, not an architecture constraint. A team running a locally hosted model on private infrastructure gets the same agent framework as a team using a commercial API. This decouples agent capability from model provider choice, which is significant for organisations with data residency or procurement constraints.
Plugin System
The 1.0 release introduces a plugin system that allows teams to extend the agent with custom tools, internal API connections, and specialist domain knowledge. Plugins can modify how the agent interprets task context, connect to internal systems not covered by a general model's training data, or add domain-specific reasoning steps to the execution loop.
SWE-Bench Performance in Context
SWE-bench Verified is a curated subset of the SWE-bench evaluation designed to remove ambiguous or low-quality tasks, making scores comparable across systems. A 68 percent result with Qwen3-Coder-480B places OpenHands among the highest performers on the leaderboard for any open-source system, matching or exceeding several commercial agentic coding products that charge on a subscription or usage basis. The 72 percent result with Claude Sonnet 4.5 and extended thinking is among the highest published scores for any autonomous coding agent to date, open or commercial.
Cost Economics: Cents Per Resolved Task
Self-hosting OpenHands with a model on H100 GPU compute costs roughly 20 US cents to US$1.05 per resolved task, depending on model selection and task complexity. For teams running significant volumes of automated engineering work — generating patches for known issues, running test suites, drafting integration code, or handling routine refactors — the per-task economics of a self-hosted open-source agent compare favourably against commercial alternatives that bundle model costs into subscription pricing. Teams with high agent usage volume see the most significant cost difference.
What OpenHands Means for Indian Engineering Teams
India's software engineering industry — from product startups to IT services firms running large developer populations — is actively evaluating autonomous coding tools. The primary constraint has been data residency: routing code and repository context to a third-party US-based API endpoint is prohibited or commercially unacceptable under many enterprise and government-sector contracts. OpenHands 1.0, with its BYOM framework and Docker sandboxing, is the most capable open-source autonomous coding agent that engineering teams can deploy entirely on domestic infrastructure. Teams can run it on AWS India or Azure India compute nodes, connect it to a locally hosted or sovereign-cloud language model, and keep all code and context within their own infrastructure perimeter. For Indian engineering teams that cannot send code to third-party cloud APIs, OpenHands 1.0 is now a credible production-grade option where previously only commercial APIs existed.
The Bottom Line
All Hands AI shipped OpenHands 1.0 on 8 September 2026. The open-source autonomous coding agent resolves 68 percent of SWE-bench Verified tasks with Qwen3-Coder-480B and 72 percent with Claude Sonnet 4.5 and extended thinking — placing it at or above commercial alternatives on the primary benchmark for autonomous software engineering. The 1.0 release delivers production-ready Docker sandboxing, a BYOM model framework, configurable resource controls and security policies, and a plugin system. Self-hosted cost runs between $0.20 and $1.05 per resolved task on H100 GPU compute. For Indian engineering teams that cannot send code to third-party cloud APIs, OpenHands 1.0 is the first open-source autonomous coding agent capable enough to act as a supervised contributor on real production engineering work.
Frequently Asked Questions
What is OpenHands 1.0 and who built it?+
OpenHands 1.0 is an open-source autonomous coding agent built and maintained by All Hands AI, a San Francisco-based company founded in 2024. The project began as OpenDevin — a community experiment in applying language model reasoning to real software engineering tasks at the level of filing and resolving GitHub issues — before All Hands AI formalised it into a production-ready platform. Version 1.0, shipped on 8 September 2026, brings production-ready Docker sandboxing, a Bring Your Own Model framework, built-in security policies, configurable resource limits, and a plugin system. The platform is available as open source under a permissive licence with no gate on the underlying agent framework.
What SWE-bench score does OpenHands 1.0 achieve?+
OpenHands 1.0 achieves 68 percent on SWE-bench Verified when paired with Qwen3-Coder-480B at up to 100 turns per task. When paired with Claude Sonnet 4.5 and extended thinking instead, the score rises to 72 percent. SWE-bench Verified is a curated subset of the SWE-bench evaluation that removes ambiguous or low-quality tasks, making results more comparable across autonomous coding systems. Both scores place OpenHands at or above several commercial autonomous coding agents on the same leaderboard, including products that charge subscription or usage-based fees for access.
How does OpenHands 1.0 handle security when executing code autonomously?+
OpenHands 1.0 runs every task inside an isolated Docker container with production-ready sandboxing. The sandbox enforces built-in security policies at the container level, with configurable resource limits on CPU, memory, and network access that administrators set per agent session. This isolation means the agent can execute code it writes — running tests, installing packages, starting processes, and verifying output — without those operations affecting the host system or other workloads. The Docker sandbox design makes isolation a first-class system guarantee rather than a best-practice recommendation, which is the key architectural difference from earlier versions of the OpenDevin project.
What does BYOM mean in OpenHands 1.0, and why does it matter for Indian teams?+
BYOM stands for Bring Your Own Model. In OpenHands 1.0, the BYOM framework lets engineering teams connect any language model API — a commercial API, a locally hosted open-weight model, or a model running in a sovereign cloud — to the OpenHands agent harness. The model choice is a configuration setting, not an architecture constraint. For Indian engineering teams, BYOM is particularly significant because many enterprise, financial services, and government-sector contracts prohibit routing code or sensitive context to third-party US-based API endpoints. BYOM means those teams can deploy OpenHands on AWS India or Azure India compute, connect it to a locally hosted model, and run the full autonomous coding workflow without any data leaving their infrastructure perimeter.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.
