Microsoft Research released the Fara1.5 family of browser computer-use agents on Friday May 22 — three sizes at 4B, 9B and 27B parameters, fine-tuned from Qwen 3.5 base. [1] On the Online-Mind2Web benchmark of 300 tasks across 136 popular sites, Fara1.5-27B scored 72 percent task success, beating OpenAI's Operator (58.3 percent), Google Gemini 2.5 Computer Use (57.3 percent) and Yutori Navigator n1 (64.7 percent). [2] Fara1.5-9B scored 86.6 percent on WebVoyager; the 4B scored 80.8 percent.

The release shipped on the same Friday the Anthropic round dominated the cap-table news cycle and the same week the paper has been tracking as Anthropic's Stainless acquisition and the SpaceX $45 billion disclosure. [1] The Fara models are released open-weight on Microsoft Foundry and HuggingFace; the recommended deployment vehicle is MagenticLite, a sandboxed Chromium harness with watch-mode action monitoring and an immediate pause control. [2] Fara1.5 is trained to halt and ask the user at three critical points — missing information, ambiguous instructions, or an irreversible action without prior approval.

The structural read is the gap between the financing story and the capability story. Microsoft shipped the best generally available computer-use model of the week and the news cycle ran a $30 billion round instead. The 27B Fara is the model an enterprise IT team would actually deploy this quarter; the round is the line in the cap table. The next test is whether either OpenAI or Google posts an Online-Mind2Web response number before Memorial Day.

-- THEO KAPLAN, San Francisco