Microsoft's Free AI Just Beat OpenAI and Google at Browsing the Web

Microsoft Research released Fara1.5, an open-weight family of browser “computer use” agent models (4B, 9B, 27B parameters) that outperform proprietary rivals on key live-web benchmarks. Fara1.5-27B scored 72% on Online-Mind2Web (vs. OpenAI Operator 58.3% and Google Gemini 57.3%) and 88.6% on WebVoyager (vs. OpenAI 87.0%). Microsoft published model weights and inference code, uses GPT-5.4 as a teacher via a FaraGen1.5 pipeline, and is already hosting the 9B variant on Azure AI Foundry. The open-source, sandboxed approach could accelerate adoption, lower barriers to on-premise deployment, and strengthen Microsoft’s competitive positioning against Google and OpenAI, with potential positive implications for MSFT’s AI product stack and Azure demand.