Microsoft Fara1.5-27B Tops Live Web Browsing Benchmarks

Microsoft Research announced Fara1.5, an open-weight browser agent family that includes 4B, 9B and 27B parameter models. The 27B variant scored 72% on the Online-Mind2Web benchmark, higher than OpenAI’s Operator at 58.3% and Google’s Gemini 2.5 Computer Use at 57.3%.

Fara1.5 models are built by fine-tuning Alibaba’s Qwen 3.5 base model. Microsoft said the 9B variant is available on Azure AI Foundry now, with the 4B and 27B models arriving shortly. All model weights and inference code are publicly released, enabling organizations to run the agent on their own hardware.

Online-Mind2Web evaluates how often an AI agent completes 300 real-world tasks across 136 live websites, including form filling, product comparison and booking. Fara1.5-27B reached 72% task success on that test. The mid-sized Fara1.5-9B scored 63.4%, also above the reported results for Operator and Gemini 2.5. On the WebVoyager benchmark, Fara1.5-27B scored 88.6%, compared with Operator at 87.0% and a 30B proprietary model, Holo2, at 83.0%.

Other reported results on Online-Mind2Web include Alibaba’s GUI-Owl-1.5 (8B) at 48.6%, AI2’s MolmoWeb at 35.3%, Microsoft’s earlier Fara-7B at 34.1%, and a proprietary system, Navigator n1, at 64.7%.

Microsoft attributed the performance gains to a rebuilt training pipeline called FaraGen1.5. The pipeline used a large external model identified by Microsoft as GPT-5.4 to generate demonstrations of browser tasks; those demonstrations formed part of the training data for Fara1.5. Researchers also created six synthetic, fully functional replicas of common websites, such as email clients and marketplaces, so the agent could practice gated or irreversible actions without using real accounts.

The release includes safety and control layers. Fara1.5 runs actions inside MagenticLite, a sandboxed browser environment that logs each step and allows users to halt the agent. The system is designed to stop and request approval before attempting actions it cannot undo. Yash Lara, senior product manager lead at Microsoft Research, noted that balancing safeguards with a smooth user experience was important and that the Magentic-UI gives people opportunities to intervene while limiting approval fatigue.

Microsoft said it plans to extend Fara1.5 beyond web browsing to desktop and enterprise software so agents can interact with a broader set of applications. The company added that public model weights and on-premise deployment options let organizations control data flow and compute costs, in contrast to cloud-only agent services.