Microsoft releases Fara1.5 models, beating OpenAI and Gemini in browser tasks

Microsoft Research's AI Frontiers lab has released the Fara1.5 series of models, designed for browser-based computer use agents. The series comes in three sizes: 4B, 9B, and 27B parameters. It works with MagenticLite, a sandbox browser interface, reading screenshots and outputting mouse and keyboard actions to complete web tasks.

Fara1.5 is built on Qwen3.5 and uses a 'observe-think-act' loop. Each step combines chat history with the last three browser screenshots to decide the next move. In tests on 136 popular websites with 300 tasks, Fara1.5-27B scored a 72% success rate, beating OpenAI Operator's 58.3% and Gemini 2.5 Computer Use's 57.3%.

The model was fine-tuned on about 2 million samples, with 60% from web traces and 12.8% from synthetic environments. For safety, it stops and asks the user when info is missing, tasks are unclear, or an irreversible action is about to happen. All actions are logged in MagenticLite, keeping a safe boundary between the agent and the user's device.