이미지: X — 인프라·칩
Summary
- Yutori's browser-operating agent model, Navigator, runs on Together AI's infrastructure
- The company says it delivers 2x faster inference and 4-5x lower costs than frontier-class models
- Browser agents work by repeating a cycle of checking screenshots and executing actions dozens of times per task
- 모델명
- Navigator (Yutori 개발)
- 구동 인프라
- Together AI
- 추론 속도
- 프론티어 모델 대비 2배 빠름
- 비용
- 프론티어 모델 대비 4~5배 저렴
- 발표 계정
- Together AI (X, 2026-08-16)
Together AI has showcased Navigator, Yutori's browser-operating model, running on its infrastructure. The company said the model delivers 2x faster inference and 4-5x lower costs 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-operating agents work in a way similar to how a person uses a website. They capture the screen to check the current state, choose a single 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," with 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 reduction in the speed and cost of each loop iteration makes a big difference in 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 understood not as Together AI having developed Navigator, but as Together AI serving a model built by Yutori on its own servers and reporting the resulting performance figures. Recently, several companies have entered the race to make browser-operating 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 real products, inference unit cost matters more than model performance benchmarks. In a structure where a single task involves dozens of rounds of screen capture and judgment, lowering the per-call cost while maintaining the same accuracy is what determines long-term service costs. If this announcement holds true, teams using Navigator can handle the same workload with fewer GPU resources — a combination worth watching for startups looking to build browser automation into commercial products.
