Microsoft’s free AI Just beat OpenAI and Google in Web browsing

Imagine telling your computer to search for vacation rentals, compare five properties, fill out a reservation form, and confirm which one is closest to the beach. You go make coffee. It’s done when you come back. That’s the promise of “computer users”—an AI that reads your browser and clicks, scrolls, and types exactly as a human would, without the need for special plugins.

OpenAI tested this first with Operatorwhich was launched in January 2025 at $200 per month before being developed into ChatGPT Agent and closing in August. Google has Gemini 2.5 Desktop Usage. They are all proprietary, cloud-based, and expensive to run.

This week, Microsoft Research released a small sample called About 1.5-and on the benchmarks they count for, they beat them both.

The family comes in three levels: 4 billion, 9 billion, and 27 billion units, all built on Qwen3.5, the version of Alibaba that Microsoft has improved the work of the browser, all weights released publicly. (Parameters are what determine the level of knowledge of an AI model, meaning it is very advanced.)

Getting there required rethinking the entire development process from the ground up. “We started with a simple question: What does it take for a small sample to be successful in aging tasks?” the AI ​​Frontiers group he wrote. “The solution involved the whole life – data generation, training objectives, model design, and calling had to be redesigned together and not in isolation.”

Measurements

Online-Mind2Web is the benchmark that is important for the work that Microsoft wanted to achieve. It measures how often an AI assistant correctly completes 300 different, real-world tasks on 136 popular websites – things like comparing products, filling out forms, and booking tasks – listed as part of the correctly completed tasks on the Internet, a variable.

Fara1.5-27B scored 72%. OpenAI users scored 58.3%. Google’s Gemini 2.5 Desktop Usage scored 57.3%. Yutori’s Navigator n1, another excellent option, reached 64.7%. Even Fara1.5-9B, the average model, hit 63.4% – ahead of OpenAI and Google.

Open opponents also failed. Alibaba’s GUI-Owl-1.5 on 8 billion units scored 48.6%. AI2’s MolmoWeb scored 35.3%. Microsoft’s previous version, Fara-7B, scored 34.1%—making this release almost double the previous version of the same size.

On WebVoyager, the second online benchmarking system did the same, Fara1.5-27B hit 88.6%, beating OpenAI Operator’s 87.0% and beating H Company’s Holo2 30 billion at 83.0%.

How it learned

The secret of the sauce is the training hose. Microsoft used a system called FaraGen1.5 to generate educational information. Here’s the clever part: they used GPT-5.4—an OpenAI model—as a “teacher assistant” to demonstrate how to complete browser tasks. The show is a tutorial for Fara1.5. You are using OpenAI’s most capable version to train an open source adversary.

They also created six fake, functional images of real websites—email clients, calendars, marketplaces—so that the model could perform tasks that required logging in or performed random actions (such as sending email or booking flights) without touching real accounts. This is called domain training, and it is a very important part of why Fara1.5 performs “gateway” tasks better than its predecessors.

Each model is designed to make you stop and ask before you do something irreversible. “Balancing strong defenses such as Key Points and seamless user journeys is essential,” Yash Lara, Senior PM Lead at Microsoft Research, he told VentureBeat. “Having a UI, such as Microsoft Research’s Magentic-UI, is important to give users the opportunity to intervene when needed, and also help prevent approval fatigue.”

This is important because OpenAI was not blind to threats when it launched the ChatGPT Agent. “If you sign a ChatGPT agent into a site or activate a connection, they can access sensitive information from those sites, such as emails, files, or account information,” the company said. he wrote.

Fara1.5 runs everything through MagenticLite, a sandboxed browser that stores everything and allows users to stop the agent at any time.

Browser AI has become a crowded competition-Google’s Gemini in Chrome, Perplexity’s Comet, Anthropic’s Claude for Chrome. The edge of Fara1.5 is that it is open source: public domain, open source GitHubit runs on the hardware you control. Fara1.5-9B is currently available Azure AI Foundry; 4B and 27B models will arrive soon. Microsoft says it plans to expand Fara1.5 beyond the browser and software for computers and businesses.