Microsoft's Fara1.5 tops web benchmarks
- Microsoft Research published the open-weight Fara1.5 browser-agent family on May 21, 2026, saying its top model outscored OpenAI and Google systems on live-web tasks. - Microsoft said Fara1.5-27B scored 72% on Online-Mind2Web, ahead of OpenAI Operator at 58.3% and Google Gemini 2.5 Computer Use at 57.3%. - Microsoft Build runs June 2-3 in San Francisco and online, where the company is expected to discuss broader agent tooling.
Microsoft Research published the Fara1.5 family of browser-focused AI agents on May 21, adding a new entrant to the contest over systems that can click, type and complete tasks on the live web. In a research post, Microsoft said its largest model, Fara1.5-27B, scored 72% on the Online-Mind2Web benchmark, ahead of OpenAI’s Operator at 58.3% and Google’s Gemini 2.5 Computer Use at 57.3%. The release covers three open-weight models — 4B, 9B and 27B — and sits inside a broader Microsoft push around agent software and smaller models that can run more efficiently. At the same time, Google said this week that the Gemini app now reaches more than 900 million monthly users, underscoring that model performance and product reach are moving on separate tracks. ### Which Microsoft system actually posted the benchmark lead? Microsoft Research said Fara1.5-27B was the top-performing model in the family on browser-use tasks. The company described Fara1.5 as a set of “computer use agent” models for the browser and named three versions: Fara1.5-4B, Fara1.5-9B and Fara1.5-27B. The 72% figure came from Online-Mind2Web, a benchmark Microsoft used to compare live-web task completion. (microsoft.com) Microsoft’s post and subsequent reports said that score exceeded OpenAI Operator’s 58.3% and Google Gemini 2.5 Computer Use’s 57.3%. ### What is Fara1.5 supposed to do in practice? Microsoft said Fara1.5 is designed for browser-based “computer use” tasks, meaning the model can act on web pages rather than only answer questions in chat. (microsoft.com) In a related Microsoft Research post, the company said Fara1.5 works with MagenticLite, an experimental application that combines browser actions, local files and model orchestration in one workflow. MagenticLite is built around what Microsoft called a system of separate components, including Fara1.5 for browser actions and another model for reasoning and delegation. Microsoft said that setup is aimed at running efficiently and keeping data on a user’s machine. ### Why does Google’s 900 million-user number matter here? (microsoft.com) Google said at its I/O 2026 keynote that the Gemini app had surpassed 900 million monthly active users, up from 400 million a year earlier. Chief Executive Sundar Pichai said Google has been focused on showing value “in the products they use every day,” tying Gemini’s growth to integration across the company’s existing services. (microsoft.com) The same keynote said Google now processes more than 3.2 quadrillion tokens per month across its surfaces, another measure of how heavily AI features are being distributed through the company’s products. Those figures do not measure browser-agent benchmark quality, but they show Google’s scale in getting AI tools in front of users. ### Is this an open release or a product launch? (blog.google) Microsoft called Fara1.5 an open-weight family, which means the model weights are being released for outside use rather than kept solely behind a hosted interface. The company framed the launch as a research release from Microsoft Research AI Frontiers, not as a standalone consumer product on the scale of ChatGPT or the Gemini app. (blog.google) That distinction matters because Microsoft and Google are showing different parts of the agent race. Microsoft is publishing model components and tooling for developers and researchers, while Google is highlighting user growth inside mass-market products such as Search and the Gemini app. That comparison is an inference from the two companies’ announcements. (microsoft.com) ### What comes next from Microsoft and Google? Microsoft Build is scheduled for June 2-3 in San Francisco and online, and the company’s event page says the conference will focus on “real code and real systems” built by Microsoft teams. That gives Microsoft an immediate venue to expand on its agent framework, model releases and developer tooling. Google’s next step is already visible in the products it announced at I/O 2026, including new Gemini app features and Gemini 3.5 Flash. (microsoft.com) The company said those updates are rolling out as Gemini expands across 230 countries and more than 70 languages. (blog.google) (build.microsoft.com)