Microsoft’s Fara1.5 AI outperforms OpenAI and Google in web tasks

Microsoft Research just dropped a set of AI models that can browse the web better than anything OpenAI or Google has built. And in a twist that should make the closed-source crowd uncomfortable, the models are open-weight.

The Fara1.5 family, released on May 22, comprises three models with 4B, 9B, and 27B parameters. The flagship 27B variant scored 72% on the Online-Mind2Web benchmark, a grueling test that throws 300 tasks across 136 live websites at browser-using AI agents. OpenAI’s Operator managed 58.3%. Google’s Gemini 2.5 Computer Use hit 57.3%. In English: Microsoft’s model completed nearly three-quarters of real-world web tasks successfully, while its Big Tech rivals couldn’t crack six in ten.

The benchmark gap is real

The 9B model scored 63.4%, which puts it ahead of both OpenAI’s and Google’s proprietary systems despite being a fraction of their size. It came within striking distance of Yutori Navigator n1, a competitive agent that posted 64.7%.

For context on how fast this space is moving: Microsoft’s previous model, Fara-7B, launched in November 2025 and scored just 34.1% on the same benchmark. That means the team roughly doubled performance in about six months.

The models are built on the Qwen3.5 architecture and use something called MagenticLite, a sandboxed browser interface that gives the agent a controlled environment to interact with web pages. They also incorporate an observe-think-act loop with a human-in-the-loop safeguard, meaning the agent pauses before executing critical actions like purchases or account changes and asks the user for confirmation.

Microsoft has made the 9B model available on Microsoft Foundry, with the 4B and 27B versions expected to follow.

Why open-weight matters here

OpenAI’s Operator and Google’s Gemini 2.5 Computer Use are proprietary systems. Fara1.5 being open-weight means developers can download, modify, and deploy these models on their own hardware. Microsoft specifically designed the Fara1.5 family to run efficiently on modest hardware, with proportional scaling benefits as you move up in model size.

The training pipeline also got a significant upgrade. Microsoft introduced FaraGen1.5, an enhanced synthetic data pipeline that generates better training examples for complex browser interactions.

What this means for crypto and DeFi

Microsoft didn’t build Fara1.5 with crypto in mind. There are no direct integrations with any blockchain protocol, DeFi application, or Web3 project.

DeFi interfaces are web applications. Swapping tokens on Uniswap, managing a vault on Aave, bridging assets across chains: these are all browser-based tasks involving forms, confirmations, and multi-step workflows — exactly the kind of thing Fara1.5 was trained to handle.

The human-in-the-loop design is particularly relevant here. DeFi transactions are irreversible. An agent that pauses before signing a transaction and asks for confirmation addresses one of the biggest risks in autonomous on-chain activity: accidentally approving a malicious contract or sending funds to the wrong address.