Microsoft Research AI Frontiers Lab released the Fara1.5 series of agent models designed for browser automation. The models, with parameter scales up to 27B, operate via the MagneticLite sandbox using an Observe-Think-Act cycle to complete web tasks. In benchmark tests, the 27B version achieved a 72% task success rate, outperforming rivals like OpenAI Operator. The models are trained on diverse data and incorporate safety features, halting for unclear tasks or privacy concerns before executing irreversible actions.