Microsoft Research has released Fara1.5, a family of open-weight browser agents that beat OpenAI's Operator and Google's Gemini 2.5 Computer Use on the industry's toughest live-web benchmark. The agents are designed to navigate and interact with real websites, automating tasks that previously required human input.
How Fara1.5 performed
The benchmark used for testing measures an agent's ability to complete complex tasks on live websites — things like booking flights, filling out forms, or comparing products. Fara1.5 scored higher than both competing models, though Microsoft Research hasn't disclosed exact figures. The company said the results show that open-weight models can compete with proprietary systems from big tech labs.
What open-weight means for developers
Unlike closed systems where the model weights remain secret, Fara1.5's weights are publicly available. That lets developers download, modify, and run the agents on their own hardware. They can fine-tune the model for specific use cases or integrate it into existing software. Microsoft Research says the goal is to speed up innovation in browser automation, a field that's seen a surge of interest as companies look to automate repetitive web tasks.
Why the benchmark matters
Live-web benchmarks are considered the gold standard because they test agents on real, changing web pages — not static datasets. A single broken link or a pop-up ad can trip up a system. Fara1.5's performance suggests it can handle these unpredictable environments better than its rivals. For businesses, that could mean fewer failed automation attempts and more reliable bot operations.
The release positions Microsoft Research as a serious contender in the race to build practical browser agents. OpenAI's Operator and Google's Gemini 2.5 Computer Use both remain accessible only through their respective cloud platforms. Fara1.5, by contrast, can be run locally or on any cloud provider. That flexibility could appeal to companies wary of vendor lock-in.
Researchers and hobbyists can already find the model weights on Microsoft's official repository. The company hasn't said when it will publish a full technical paper or release a larger version of the agent.