The OpenAGI Foundation has launched Lux, a new computer use model that aims to automate tasks across desktops and web browsers. Lux is designed to take natural language instructions, interpret them, and perform actions like clicks and keystrokes on various applications. It has achieved an impressive score of 83.6 on the Online Mind2Web benchmark, which evaluates over 300 real-world tasks. This score places Lux ahead of competitors, including Google Gemini CUA, OpenAI Operator, and Anthropic Claude Sonnet 4.
Lux is not just another chat model; it operates directly on rendered user interfaces rather than relying on specific application programming interfaces (APIs). This means it can handle a wide range of desktop applications, from web browsers to email clients and spreadsheets. The model is accessible to developers through the OpenAGI SDK and API console.
The team behind Lux has identified several key use cases, including software quality assurance, deep research tasks, social media management, online store operations, and bulk data entry. In these scenarios, Lux can perform numerous user interface actions while staying aligned with the instructions given in natural language.
Lux offers three execution modes tailored to different needs. The Actor mode is the quickest, taking about one second per step for straightforward tasks. The Thinker mode is more suited for complex goals, breaking them down into smaller tasks. Finally, the Tasker mode allows for precise control, where users can provide a detailed list of steps for Lux to execute.
In terms of performance, Lux completes tasks significantly faster than its competitors, averaging around one second per step compared to OpenAI Operator’s three seconds. Additionally, Lux is reported to be ten times cheaper per token than its rivals, making it a cost-effective solution for long-running tasks.
A unique aspect of Lux’s training is its method called Agentic Active Pre-training. This approach allows Lux to learn from interacting with digital environments, rather than just passively consuming text. The OpenAGI team has also developed OSGym, an open-source platform that can run multiple operating system replicas simultaneously to support the training of models like Lux.
Overall, Lux represents a significant step forward in automating computer tasks, offering a powerful tool for developers and businesses looking to streamline operations. The OpenAGI Foundation’s commitment to open-source development through OSGym further enhances its potential for widespread use and innovation in the field of artificial intelligence.