09:01 ET OpenAGI Releases Lux, the Most Performant Computer-Use Model

Lux sets a new standard for computer use and comes with an SDK to empower developers to build real-world computer-use applications , /PRNewswire/ -- OpenAGI Foundation, a computer-use research lab, emerged from stealth today with the release of Lux, its frontier foundation model. Lux achieved a state-of-the-art score of 83.6 on the Online-Mind2Web benchmark, a collection of over 300 real-world computer-use tasks. This outperforms models from major AI labs, including Google's Gemini CUA (69.0), OpenAI's Operator (61.3), and Anthropic's Claude Sonnet 4 (61.0). Built by top AI researchers from MIT, Carnegie Mellon University, and the University of Illinois Urbana-Champaign, Lux delivers more than higher accuracy. The model is faster (1 second per step vs. 3 seconds for OpenAI Operator) and more cost-efficient (10 times cheaper per token). Lux comes with an SDK to empower developers to build computer-use applications such as automating software QA, browser-based deep research, social media management, online store management, data entry, and bulk operations, among many others. Lux can operate across desktop applications and is not confined to the browser like ChatGPT Atlas. Lux works in three modes: Actor for quick tasks, Thinker for multi-step goals, and Tasker, which gives the user more control by accepting a list of "todo"s that break down the task it must complete. "For 60 years, computers have been just tools. Now, with agentic computer-use technology, they are becoming intelligent entities that understand our intentions and navigate complexity on our behalf," said Zengyi Qin, founder and CEO of OpenAGI Foundation. "At OpenAGI, we are taking a step towards that vision, making computers reliable companions to humans." OpenAGI trained Lux with a novel technique called Agentic Active Pre-training, using a large volume of computer-use trajectories and action sequences. This allows the model to produce actions that control the computer and respond to real-world tasks. With this release, OpenAGI is also open-sourcing OSGym, the data engine and infrastructure for training generalizable computer-use agents. Press contact: OpenAGI press center[email protected] SOURCE OpenAGI Foundation