Microsoft's Free AI Just Beat OpenAI and Google at Browsing the Web
Microsoft Research unveiled Fara1.5, open-weight AI models that beat OpenAI Operator and Google Gemini 2.5 on browser-based tasks. The 27B model scored 72% on Online-Mind2Web vs. 58.3% and 57.3%. Built on Qwen 3.5 and trained with synthetic data, they're free on Azure AI Foundry.
Quick Take
Fara1.5-27B scores 72% on Online-Mind2Web vs OpenAI’s 58.3% and Google’s 57.3%.
Models are open-weight, 4B/9B/27B parameters, fine-tuned Qwen 3.5.
Trained using GPT-5.4 as teacher and synthetic domain replicas for gated tasks.
Free access on Azure AI Foundry, outperforming proprietary rivals.
Market Impact Analysis
NeutralOpen-source AI advancement may foster innovation in decentralized AI applications, but direct crypto market impact is minimal.
Speculation Analysis
Key Takeaways
- Fara1.5-27B scored 72% on Online‑Mind2Web, crushing OpenAI Operator’s 58.3% and Google Gemini 2.5’s 57.3%.
- The open‑weight family ships in 4B, 9B, and 27B parameter sizes, all fine‑tuned from Qwen 3.5.
- Training used GPT‑5.4 as a teacher with synthetic domain replicas to master gated web tasks.
- Free on Azure AI Foundry, the release pressures proprietary computer‑use agents in cost and access.
What Happened
Microsoft Research dropped Fara1.5, an open‑weight AI model family that outperforms proprietary rivals at browser‑based tasks. The 27‑billion‑parameter version scored 72% on Online‑Mind2Web, a live‑web benchmark of 300 real‑world assignments like comparing products and booking services. That leaves OpenAI Operator (58.3%) and Google Gemini 2.5 Computer Use (57.3%) in the dust. Even the mid‑sized 9B model hit 63.4%, staying ahead of both tech giants. Built on fine‑tuned Qwen 3.5, the entire lineup is freely available on Azure AI Foundry, handing developers a production‑grade computer‑use agent at zero cost.
The Numbers
On the industry‑standard Online‑Mind2Web test, Fara1.5‑27B’s 72% success rate marks a double‑digit leap over top closed models. For context, Yutori’s Navigator n1, the leading proprietary alternative, managed 64.7%. Open‑source contenders fared worse: Alibaba’s GUI‑Owl‑1.5 (48.6%) and AI2’s MolmoWeb (35.3%). On the WebVoyager benchmark, Fara1.5‑27B reached 88.6%, edging past Operator’s 87.0% and H Company’s Holo2 at 83.0%. Microsoft’s own earlier model, Fara‑7B, captured just 34.1%, meaning this generation nearly doubles performance at a comparable size.
Why It Happened
The breakthrough stems from a new training pipeline called FaraGen1.5. Microsoft researchers used GPT‑5.4 as a teacher model and generated synthetic data that mimics gated, real‑world browser environments. This allowed the team to fine‑tune Qwen 3.5 into a specialist that reads screens, clicks, scrolls, and types without custom APIs—all while keeping weights open. The team redesigned data generation, objectives, and model design together, rather than tweaking each in isolation. The result is a compact model that generalises across live websites, not just static test pages.
Broader Impact
Fara1.5 tilts the field toward open‑source tooling for agentic tasks. A free, best‑in‑class computer‑use model lowers barriers for startups and researchers, potentially sparking a wave of browser‑automation apps. It also challenges the economics of proprietary agents, which often require cloud subscriptions. As more efficiency‑focused models appear, demand could shift away from bloated, general‑purpose giants toward smaller, task‑specific open weights.
What to Watch Next
- Developer uptake on Azure AI Foundry and integration into productivity tools.
- Whether OpenAI or Google counter with cheaper or open models for browser tasks.
- Community‑built fine‑tunes extending Fara1.5 into e‑commerce, travel, or data entry.
This article is for informational purposes only and does not constitute financial advice.
Always late to trends?
Join for the latest news, insights & more.
Disclaimer: Bytewit is an independent media outlet that delivers news, research, and data.
© 2026 Bytewit. All Rights Reserved. This article is for informational purposes only.
