Fara-1.5: Scalable Learning Environments for Computer Use Agents

Fara-1.5 introduces a scalable data pipeline, FaraGen1.5, designed for training computer use agents (CUAs) through modular components: environments, solvers, and verifiers. The pipeline utilizes both live and synthetic environments, powered by models like GPT-5.4, and employs a supervised finetuning approach to produce three variants of Fara1.5 (4B, 9B, and 27B). The 9B and 27B models achieve state-of-the-art performance on browser-use benchmarks, with scores of 63.4% and 72.3% on Online-Mind2Web, respectively, demonstrating significant advancements in efficiency and task correctness for practitioners developing LLM-based agents.