Microsoft's Fara1.5 beats rivals

Microsoft Research released Fara1.5 on May 21 as a family of open-weight browser agents designed to carry out tasks on live websites, adding a new entrant to the market for AI systems that can click, type and navigate on a user’s behalf. Microsoft said the models come in 4 billion, 9 billion and 27 billion parameter versions and are part of a broader package that also includes MagenticLite, an experimental agentic application, and MagenticBrain, a reasoning and orchestration model. In its research write-up, Microsoft said Fara1.5 was built for browser-based tasks such as form filling and cross-site comparison shopping. Decrypt reported on May 23 that Microsoft’s release outperformed OpenAI’s Operator and Google’s Gemini 2.5 Computer Use on benchmark tests. ### What exactly did Microsoft release? Microsoft Research said Fara1.5 is an “open-weight” family, meaning the model weights are available for use rather than being kept fully proprietary. The company described the system as a computer-use agent for browser tasks and said it was optimized to work with smaller models running more efficiently on user hardware. (microsoft.com) The May 21 Microsoft post said MagenticLite combines browser and local file-system workflows in one system. Microsoft said MagenticBrain handles reasoning, delegation and terminal use, while Fara1.5 handles browser interaction. ### Which benchmark result is drawing attention? Microsoft said Fara1.5-27B scored 72% on Online-Mind2Web, a benchmark made up of 300 tasks across 136 websites. (microsoft.com) The company also said Fara1.5-9B reached 63% and Fara1.5-4B reached 57%, compared with 49% for GUI-Owl-1.5-8B, which it identified as the prior best model at that scale. Decrypt reported that those results put Fara1.5 ahead of OpenAI’s Operator and Google’s Gemini 2.5 Computer Use in Microsoft’s comparison table. (microsoft.com) Microsoft said the 27B model also narrowed the gap with Yutori Navigator n1, another browser agent in the field. ### Why does the “open-weight” label matter here? Microsoft’s decision to publish open weights puts Fara1.5 in a different category from closed commercial browser agents such as Operator. (microsoft.com) That matters because developers can inspect, adapt and run the models in their own environments, subject to Microsoft’s licensing terms and deployment options. (decrypt.co) Decrypt said the release highlighted how open models are becoming more competitive with proprietary systems in web automation. That framing came as Microsoft emphasized lower-cost operation and local execution as part of the product design. ### How much of this is about speed, not just accuracy? Microsoft said Fara1.5 was trained for “realistic interactions” and tied the work to live browsing tasks people actually perform. (microsoft.com) Decrypt said benchmark coverage included latency measures in real-world browsing tests, not only final task completion. The company’s research blog argued that agent performance depends on tool orchestration and action rather than knowledge alone. (decrypt.co) Microsoft said that approach lets smaller models handle a broader range of tasks at lower cost. ### Where does this leave Microsoft in the OpenAI story? Yahoo Finance reported on May 23 that some analysts see Microsoft as a beneficiary if OpenAI moves ahead with an initial public offering, because of Microsoft’s strategic stake and broader AI position. (microsoft.com) The report said Microsoft stock looked attractive ahead of any OpenAI listing. (microsoft.com) Microsoft’s Fara1.5 release does not depend on OpenAI, but it adds to the company’s case that it can compete across multiple layers of the AI stack, from infrastructure to models to end-user agents. That is an inference from Microsoft’s product lineup and the analyst view reported by Yahoo Finance. ### What happens next? (finance.yahoo.com) Microsoft said the Fara1.5 family is part of its MagenticLite system, and third-party reports citing Azure AI Foundry listings said the 9B model was already available there, with 4B and 27B versions expected later. Microsoft’s own research pages point developers to model documentation and benchmark tables for the next step in evaluating deployment. (binance.com) (finance.yahoo.com)