Microsoft’s Fara1.5 Outperforms OpenAI and Google in AI Browser Tasks

Microsoft's latest AI model, Fara1.5, has scored an impressive 72% on the Online-Mind2Web benchmark, surpassing both OpenAI's Operator and Google's Gemini 2.5. This performance breakthrough signals a shift in the abilities of AI agents designed for web browsing tasks, with Fara1.5 emerging as a frontrunner thanks to its innovative open-source approach.

The Fara1.5 model family consists of three sizes—4 billion, 9 billion, and 27 billion parameters—all built on the fine-tuned Qwen 3.5. Microsoft has made the weights of these models publicly accessible, setting them apart from the proprietary offerings of competitors like OpenAI and Google. OpenAI's Operator, launched in early 2025, achieved only 58.3%, while Google’s Gemini 2.5 scored 57.3%. Fara1.5's performance marks a significant advancement for Microsoft in the AI field.

A New Paradigm for AI Agents

The Online-Mind2Web benchmark evaluates the effectiveness of AI agents in completing over 300 diverse tasks across popular websites, simulating real-world scenarios such as product comparisons and form submissions. Fara1.5 not only outperforms its proprietary rivals but also demonstrates a notable improvement over its predecessor, Fara-7B, which managed just 34.1% on similar tasks.

The development of Fara1.5 underwent a complete reimagining. Microsoft’s AI Frontiers team concentrated on integrating various aspects of AI development, from data generation to model orchestration, into a unified training pipeline. This approach is encapsulated in their guiding question: "What does it take to make a small model genuinely good at agentic tasks?" This inquiry led to the creation of training data using OpenAI's own GPT-5.4 as a 'teacher agent', establishing a new standard for AI model training.

Enhancing User Experience and Safety

Alongside the launch of Fara1.5, Microsoft has introduced advanced safety features. The model is programmed to pause and seek user confirmation before carrying out irreversible actions, such as sending emails or making bookings. Yash Lara, Senior PM Lead at Microsoft Research, highlighted the need to balance user safety with functionality, stating, "Every model is designed to stop and ask before doing something it cannot undo."

Fara1.5 operates within a sandboxed environment called MagenticLite, which logs all actions and allows users to intervene at any time. This approach contrasts sharply with OpenAI's ChatGPT Agent, which has raised concerns about data security when integrated with third-party platforms. Microsoft’s emphasis on user control and transparency is a key consideration in the competitive AI market.

Implications for the Future of AI

As the market for AI agents evolves, Fara1.5's open-source model may set new expectations for accessibility and performance. The model is currently available on Azure AI Foundry, with plans to release the remaining sizes soon. Microsoft also aims to expand beyond browser tasks into desktop and enterprise software applications.

The competitive landscape includes not only giants like Google and OpenAI but also emerging players such as Anthropic’s Claude and Perplexity’s Comet. However, Fara1.5’s open model offers a distinct advantage, enabling developers and users to customize and deploy the AI on their own hardware. This could pave the way for a surge of innovative applications across various domains, from personal assistance to enterprise solutions.

The introduction of Fara1.5 signifies a crucial moment in the AI agent sector, demonstrating that the blend of openness and high performance can challenge established norms and promote more inclusive development practices. As Microsoft continues to advance its AI technologies, the implications for both users and developers will be profound, potentially transforming the future of human-computer interaction.