Microsoft has launched the open-source Fara1.5 web agent model, which outperforms OpenAI and Google in key benchmarks. The Fara1.5-27B model achieved a 72% task completion rate on the Online-Mind2Web benchmark, surpassing OpenAI Operator's 58.3% and Google's Gemini 2.5 at 57.3%. Additionally, in the WebVoyager test, Fara1.5-27B scored 88.6%, edging out OpenAI Operator's 87.0%. Fara1.5 is available in three versions—4B, 9B, and 27B—fine-tuned from Alibaba's Qwen3.5 base model. Microsoft has released these models with open weights and inference code, allowing developers to deploy them on their own hardware. The 9B version is currently available on Azure AI Foundry, with the other versions to follow. The model's improvements are attributed to a new training process using FaraGen1.5, which employs GPT-5.4 as a "teacher agent" to generate training data. Microsoft has also implemented security measures, including manual confirmation for irreversible actions and a sandboxed environment for safe execution.