OpenAGI Foundation Launches Lux: A Foundation Computer Access Model for Online Mind2Web with Osgym at Scale

How do you slowly turn, manually click and work across all browsers and desktops into a reliable, automated system that can really run the computer for you at scale? Lux is the latest example of computing agents moving from research demos to infrastructure. The Operagi Foundation team is out It helpedthe base model runs on real desktops and browsers and reports 83.6 points on the online Internet benchmark, which includes 300 real computer tasks. This is ahead of Google Gemini Cua at 69.0, Opelai operator at 61.3 and anthropic claude sonnet 4 at 61.0.

What Is Lux Actually?

Lux is a computing model, not a browser plugin conversational model. It takes natural language processing, screen views, and outputs low-level actions such as clicks, key presses and hover events. It can run browsers, editors, spreadsheets, email clients and other desktop applications because they work on a given UI, not on a specific API.

From a developer's point of view, Lux is available with OpenAgi SDK and API Console. The Research Team defines targeted workloads that include software QA workflows, deep dive runs, social media management, online store operations and bulk data entry. In all these settings the agent needs to sequence dozens or hundreds of UI actions while staying aligned with natural language interpretation.

Three methods of execution for different levels of control

Lux ships with three ways of execution That reveals various trade-offs between speed, autonomy and control.

Actor mode the fastest way. It runs at 1 second per step and is intended for clearly defined tasks such as filling out a form, pulling a report from a dashboard or extracting a small set of fields from a page. Think of it as a low latency macro engine that already understands natural language.

Thinking mode Handles specific or multiple objectives. It clears the higher level instruction in the lower sub-tasks and removes them. Example Tasks include Multi Page research, TRECH for long e-mail lines or navigation of analysis sites where the exact path is defined.

Employee mode he gives a great deal. The caller provides a Python-specific list of steps that exit each Lux and return until the sequence completes or hits a hard failure. This allows teams to keep task graphs, guardrails and failover policies in their code while transferring UI control to the model.

Work, play and thinking are the three main ways of working in a process, to be done quickly and to solve complex objectives.

Benchmarks, latencies and costs

In Online Mind2Web, Lux achieves a success rate of 83.6 percent. The same bench reports 69.0 percent of Gemini Cua, 61.3 percent of user 61.0 percent of Claude Sonnet 4. So Claude Sonnet 4. So it is a useful representative for active agents running browsers and web applications.

Latency and cost are where the numbers become important to engineering teams. The OpenAgi team reports that Lux completes each step in about 1 second, while the Opelai operator takes about 3 seconds for the same step in the same test. The research group also mentions that Lux is about 10 times cheaper per token than the operator. For any agent that can easily handle hundreds of steps in a session, these constants determine whether the load is running in production.

Effective pre-training agentic and why osgym is important?

Lux is trained in a way that the research group calls for Active pre-training agent. The group combats this with a standard language training program that only includes text from the Internet. The idea is that Lix learns from working in digital environments and refines its performance through greater communication, rather than reducing the loss of predicting tokens in grassroots logs. The utility goal is different from classical reinforcement learning, and is set higher to favor exploration and understanding rather than manual reward.

This training setup relies on a data engine that can generate multiple active workstations in parallel. The OpenAgi team has already unlocked that engine as gym, Under the MIT license which allows for both research and commercial use. Osgym works with the full functionality of the operating system, not only browser boxes, and supports the functions of office software, browsers, development tools and many workflows.

Key acquisition

1. Lux is a basic computing model that runs full desktops and browsers and reached a success rate of 83.6 percent on the Internet bench of Mind2Web, ahead of gemini cua, operaite operator and claude sonnet-4.
2. Lux presents 3 modes, Actor, thinker and workerwhich includes latency UI macros, multi-step intent decomposition and intent script execution that determine the production workflow.
3. Lux is reported to have advanced by 1 second per step and is 10 times cheaper on the other side per token than Opelai, which is more important in the prison of agents who go through the actions of each job.
4. Lux is trained with an active pre-training agent, where the model learns by working in environments, rather than consuming only static web text, which aims at robustness of the screen in pure language behavior.
5. Osgym, the open source engine behind lux, can run more than 1,000 os replicas and produce more than 1,400 tractiories per minute at low cost for training and testing its active agents.

Look Official announcement, project and It's a waste. Feel free to take a look at ours GitHub page for tutorials, code and notebooks. Also, feel free to follow us Kind of stubborn and don't forget to join ours 100K + ML Subreddit and sign up Our newsletter. Wait! Do you telegraph? Now you can join us by telegraph.

Michal Sutter is a data scientist with a Master of Science in Data Science from the University of PADOVA. With a strong foundation in statistical analysis, machine learning, and data engineering, Mikhali excels at turning complex data into actionable findings.