OpenAGI comes out of stealth with Lux, an agent built to run your computer

Big tech spent the last year convincing us that AI agents would soon take the steering wheel and drive our computers. In practice, most of them still trip over dynamic pop-ups and clunky checkout forms.

Now, a stealth startup out of San Francisco is claiming it has cracked the code. OpenAGI, founded by MIT researcher and CEO Zengyi Qin, just stepped into the spotlight with Lux, a foundation model specifically built to operate computers by interpreting screen captures and firing off mouse clicks and keystrokes.

OpenAGI claims Lux achieved an 83.6% success rate on Online-Mind2Web. For context, OpenAI's Operator scored 61.3% on the same benchmark, while Anthropic's Claude Computer Use trailed at 56.3%.

Surviving the messy live web

Online-Mind2Web has quickly become the benchmark researchers take seriously because it avoids the traps of older tests. Developed by researchers at Ohio State University and UC Berkeley, and accepted at the Conference on Language Modeling 2025, it throws agents into live, dynamic websites rather than sanitized, cached pages.

The benchmark consists of 300 tasks spread across 136 real-world sites, covering everything from booking flights to navigating multi-step e-commerce checkouts. When the Ohio State team tested existing commercial agents, they found most marketing hype fell flat: many well-funded agents failed to beat SeeAct, a comparatively simple agent released in January 2024.

Even OpenAI's Operator only hit 61% in human evaluations, leading researchers to warn that practical autonomy was farther away than press releases suggested.

Training on clicks instead of tokens

OpenAGI says its edge comes from ditching standard LLM recipes. Most frontier models spend their pre-training predicting the next text token in a massive text corpus. They know how to write essays, not navigate a graphical user interface.

Lux was built using what OpenAGI calls "agentic active pre-training." Instead of pure text, it was trained directly on desktop screenshots paired with action sequences. The model learns visual layout comprehension alongside corresponding inputs like mouse clicks, typing, and navigation commands.

Qin told VentureBeat that this setup creates a self-reinforcing loop. When the model takes actions, it actively explores the software environment, generates new observations, and feeds that fresh experience back into the training pipeline. That loop could explain how a smaller, capital-constrained startup claims to outmaneuver the balance sheets of OpenAI and Anthropic.

Why it matters

For power users and developers tired of brittle browser automation scripts, an agent hitting an 80-plus percent success rate on live websites changes the equation. If Lux holds up under daily use, mundane chores like cross-platform data entry and multi-site bookings could finally become dependable enough to delegate without constant babysitting.

For the AI landscape, OpenAGI's pitch poses a fundamental question: are general-purpose chatbots retrofitted with screen-viewing tools the wrong bet for computer control? If training native action models directly on interface states delivers a 20-point leap over OpenAI at a fraction of the inference cost, the agent race might shift toward specialized, action-first architectures.

We think OpenAGI's active exploration loop is a genuinely clever way to bypass the web data ceiling. That said, claiming benchmark victory is one thing; navigating enterprise permission walls, desktop edge cases, and high-stakes workflows without disastrous mistakes is another. It looks promising, but we will need to see independent third parties put Lux through its paces before crowning a new king of computer control.