Performance and Architecture

Microsoft Research has introduced Fara1.5, a family of computer-use agent (CUA) models available in 4B, 9B, and 27B parameter sizes. Built on Qwen3.5 base checkpoints, these models function as pixel-to-action agents that interpret browser screenshots to execute mouse and keyboard inputs. On the Online-Mind2Web benchmark, the Fara1.5-27B model achieved a 72% task success rate, significantly outperforming OpenAI’s Operator (58.3%) and Gemini 2.5 Computer Use (57.3%).

The agents operate via an observe-think-act loop, processing conversation history alongside the three most recent browser screenshots. Beyond standard navigation, Fara1.5 includes meta-actions for context management, such as memorizing facts and querying users for clarification, which enables the agents to handle longer task horizons and collaborative workflows.

The FaraGen1.5 Synthetic Pipeline

The performance gains are largely attributed to FaraGen1.5, a synthetic data pipeline consisting of environments, solvers, and verifiers. To handle gated domains (e.g., email or calendar tasks requiring authentication), the team created six 'FaraEnvs'—synthetic clones with functional APIs and persona-based seed data. This allows for rigorous testing where the ground truth is known. A solver agent, powered by GPT-5.4, generates training trajectories that are then filtered by three verifiers: correctness (rubric-based), efficiency (penalizing redundant actions), and user-interaction verification (ensuring critical pauses). This pipeline was trained on approximately two million samples, with 60% consisting of web trajectories.

Safety and Deployment

Fara1.5 is designed with built-in safety triggers. The agent is trained to pause and request user input if a task requires sensitive personal information, if the instructions are ambiguous, or if an irreversible action is about to be executed. The models are integrated with MagenticLite, a sandboxed browser interface that provides a security boundary between the agent and the host machine, while ensuring all actions are logged and auditable.