Microsoft Launches New Fara1.5 Series AI Agent Model with a Task Success Rate of 72%! Exceeding OpenAI!

Microsoft Research's AI Frontiers Lab has recently introduced the new Fara1.5 series of AI agent models, designed specifically for browser interaction to enhance the intelligence of computer use. The series includes three variants with 4B, 9B, and 27B parameters, aiming to transform the browser into a more perceptive and capable intelligent assistant.

Integrating seamlessly with the MagneticLite sandboxed browser interface, Fara1.5 can directly read browser screenshots and automate web-based tasks by simulating mouse and keyboard actions. Its core operation follows a "See-Think-Act" loop, where each step combines conversation history with the three most recent screenshots to generate reasoning and subsequent actions.

In terms of performance, the Fara1.5-27B model achieved a notable 72% task success rate on the Online-Mind2Web benchmark, surpassing competitors like OpenAI's Operator (58.3%) and Google's Gemini 2.5 Computer Use (57.3%). Even the more compact 9B variant reached a 63.4% success rate, highlighting the series' efficiency.

The model was fine-tuned on a diverse dataset of approximately 2 million samples, including web trajectories (60%), synthetic environments (12.8%), user interactions (12.5%), factually-grounded data (8.8%), and visual question answering (4.9%). This multi-faceted training approach significantly improved the model's adaptability and intelligence.

To ensure user safety and privacy, Fara1.5 is designed to pause and request user confirmation in three key scenarios: when personal information is missing, when task instructions are unclear, and before executing any unapproved irreversible actions. All operations are logged within the secure MagneticLite sandbox, maintaining a clear safety boundary between the agent and the user's device.

The Fara1.5 series marks a step forward in making AI a practical and secure part of the daily browsing experience, showcasing the significant potential of AI in everyday computing.

Key Highlights:

▸ Achieves a leading 72% task success rate, outperforming key competitors.

▸ Operates on an intuitive "See-Think-Act" loop for human-like interaction.

▸ Implements robust safety measures with user confirmation prompts and secure sandboxing.