Microsoft's Fara1.5 AI model outperforms OpenAI and Google in web browsing tasks, achieving a 72% score in benchmark tests.
Summary
- Fara1.5-27B scored 72% on the Online-Mind2Web benchmark, outperforming OpenAI's 58.3% and Google's 57.3%.
- The model is available in 4 billion, 9 billion, and 27 billion parameter versions, with the 9 billion version currently on Azure AI Foundry.
- Microsoft aims to extend Fara1.5's functionality beyond web tasks in future updates.
Microsoft has introduced its new AI model, Fara1.5, which has demonstrated superior performance in web browsing tasks compared to competitors OpenAI and Google. Fara1.5-27B achieved a remarkable score of 72% on the Online-Mind2Web benchmark, significantly outpacing OpenAI's Operator at 58.3% and Gemini 2.5 at 57.3%. The Fara1.5 series includes models with 4 billion, 9 billion, and 27 billion parameters, all based on the optimized Qwen 3.5 model.
This innovative AI aims to revolutionize how users interact with the internet by functioning as a computer use agent capable of executing tasks like booking vacations autonomously. Microsoft has made the model's weights publicly available, a move that emphasizes its commitment to open-source technology and accessibility.
To achieve its high performance, Microsoft revamped its development approach. The AI Frontiers team stated, "We began with a fundamental question: What is required for a small model to excel in agentic tasks?" They focused on a collective rethinking of data generation and training objectives.
Benchmark Results
The Online-Mind2Web benchmark evaluates AI agents on their ability to complete a variety of tasks across popular websites. Fara1.5-27B's score of 72% indicates a strong capability in tackling real-world challenges, whereas the 9 billion parameter version scored 63.4%, also higher than both OpenAI and Google.
In comparison, other models like Alibaba's GUI-Owl-1.5 and AI2's MolmoWeb scored significantly lower, with 48.6% and 35.3%, respectively. This showcases the competitiveness of Fara1.5 within the AI landscape.
Moreover, in the WebVoyager benchmark, which assesses task completion on the live web, Fara1.5-27B achieved 88.6%, leading over OpenAI's 87.0% and surpassing Holo2's 83.0% performance.
Training Methodology
The success of Fara1.5 can be attributed to its innovative training methodology, which included using OpenAI's GPT-5.4 as a "teacher agent" to generate training data. This approach allowed Microsoft to leverage a competitor's technology to enhance its own model. Additionally, the creation of simulated replicas of various websites enabled the model to practice complex tasks without real-world ramifications.
Fara1.5 incorporates user-safety features, requiring confirmation before executing irreversible actions. Yash Lara, Senior PM Lead at Microsoft Research, emphasized the importance of balancing user safety with a smooth operational experience. The model operates within a secure environment, MagenticLite, ensuring users can monitor and control actions at any time.
As competition in the AI browsing space heats up, with offerings from Google and others, Fara1.5's open-source nature and adaptability position it favorably for future developments. Microsoft plans to enhance Fara1.5's capabilities to encompass desktop and enterprise applications moving forward.
