Yutori's Browser Agent Runs Twice as Fast on Together AI

Together AI has introduced Navigator, Yutori's browser-control model now running on its infrastructure, saying it delivers 2x faster inference and 4-5x lower cost compared to frontier-class models. A post from Together AI's account on X (formerly Twitter) included footage of the agent automatically operating a browser screen.

Browser-control agents work in a way similar to how humans use websites. They capture the screen to check the current state, choose one action such as clicking a button or entering text, then capture the screen again to decide the next action. Together AI described this process as a short loop of "screenshot, action, repeat" — the key point being that this loop runs dozens of times to complete a single task. Since the entire sequence of scanning search results, filling out forms, and moving to the next page happens within this loop, even a small improvement in the speed and cost of a single loop iteration can significantly change the total task time and cost.

Together AI is not a company that trains its own models but an inference cloud that runs open models like Llama and Qwen as well as partner models like Navigator on behalf of others. This announcement should be read not as Together AI having developed Navigator, but as Together AI serving a model built by Yutori on its own servers and reporting performance figures for it. Recently, several companies have entered the race to make browser-control agents lighter and faster, including Cloudflare's agent-dedicated browser Kitesurf and Nous Research's Hermes browser-use mode.

For teams trying to integrate browser agents into actual services, inference unit cost matters more than benchmark scores. In a structure where a single task involves dozens of rounds of screen capture and decision-making, lowering the cost per call while maintaining the same accuracy is what determines service costs in the long run. If the announcement holds true, teams using Navigator could handle the same workload with fewer GPU resources — a combination worth watching for startups looking to build browser automation into commercial products.