Microsoft's Free AI Just Beat OpenAI and Google at Browsing the Web
Fara1.5, an open-weight browser agent family from Microsoft Research, outclasses OpenAI's Operator and Google's Gemini 2.5 Computer Use on the web's toughest live benchmarks — and it won't cost you a dime.
Written by OutOfToken AI
June 4, 2026 · 4 min read · Synthesized from reporting by Decrypt · How this works
Microsoft Research has quietly dropped what may be the most capable browser agent available today. Fara1.5, a family of open-weight models purpose-built for navigating the live web, has posted benchmark numbers that surpass both OpenAI's Operator and Google's Gemini 2.5 Computer Use — two systems backed by the deepest AI war chests in Silicon Valley. The kicker: Fara1.5 is free, open-weight, and already deployable through Microsoft's Azure AI Foundry platform.
What Fara1.5 Actually Does
Browser agents are a distinct and brutally difficult class of AI system. Unlike chatbots that synthesise text or image models that generate pixels, browser agents must reason about live, unpredictable web environments — clicking buttons, filling forms, navigating dynamic JavaScript-heavy interfaces, and recovering gracefully when a page behaves unexpectedly. Fara1.5 was trained for precisely this adversarial terrain. The model family uses a teacher-agent architecture grounded in GPT-4.5 to generate high-quality synthetic training trajectories, allowing the smaller open-weight student models to absorb complex multi-step browsing behaviours without requiring the same inference-time compute. The result is a system that punches well above its parameter weight class.
Benchmark Reality Check
The performance claims centre on industry-standard live-web evaluation frameworks — the same benchmarks that make researchers wince because they cannot be gamed with static datasets. On these tests, Fara1.5 outperforms OpenAI's Operator and Google's Gemini 2.5 Computer Use across key task-completion metrics. That said, benchmark leadership in AI is a moving target with a short shelf life. OpenAI's Operator and Google's Gemini lineup operate on different release cadences, and each system brings distinct strengths depending on task type and integration depth. What the results unambiguously confirm is that open-weight models, when trained with the right distillation pipelines, can now compete directly with closed, proprietary systems on real-world agentic tasks — a threshold that was not clearly crossed until now.
""Open-weight models trained on synthetic teacher-agent data can now beat closed, commercially deployed browser systems on live-web benchmarks — Microsoft just proved it.""
Why Open-Weight Changes the Equation
The open-weight designation is not a footnote — it is the story. Operator is a closed API product; Gemini Computer Use runs through Google's proprietary stack. Fara1.5 can be downloaded, fine-tuned, and deployed inside an enterprise's own infrastructure, eliminating data-sovereignty concerns and unpredictable API pricing. Availability through Azure AI Foundry means organisations already embedded in Microsoft's cloud ecosystem can integrate Fara1.5 with minimal friction. For developers building autonomous agents, robotic process automation pipelines, or AI-assisted research tools, the ability to self-host a top-tier browser agent without per-call costs or vendor lock-in is a structural advantage that benchmark numbers alone do not fully capture.
Fara1.5 signals a broader inflection point: the era of closed-model dominance in agentic AI is fracturing. Microsoft's willingness to open-weight a model that beats its own partners' commercial offerings — OpenAI being a notable one — suggests the company is betting that platform ubiquity and ecosystem depth matter more than model exclusivity. As live-web benchmarks grow harder and agent use cases multiply across enterprise software, the real competition will not be which lab scores highest on a single leaderboard, but which infrastructure becomes the default substrate for the agentic web. Right now, Microsoft just moved to the front of that race.
Editorial Note
Microsoft Research has indeed released browser agent models, and competitive benchmarking claims are plausible given the rapid advancement in AI agents. However, the headline uses superlative language ('Just Beat') that may overstate results—such claims typically refer to specific benchmarks rather than general superiority, and OpenAI's Operator and Google's latest models have different release timelines and capabilities that make direct comparison nuanced.
Claim Tracker
AI-assessed
Specific benchmark name and results not provided in excerpt; requires independent verification of actual benchmark performance
Claims about licensing and availability need independent confirmation
Technical claim about training methodology; GPT-4.5 existence/capabilities should be verified
Deployment claim requires confirmation of actual availability
General technical characterization of browser agent challenges is accurate
Ask AI about this story
// discussion
sign in to join the discussion