This week, we feature insights from Radical Ventures portfolio company Yutori on their recent release of Navigator, a state-of-the-art vision-based AI agent that autonomously completes web tasks. Navigator outperforms frontier models by 10-20% in accuracy across benchmarks, operates 2-3x faster, and is uniformly preferred in head-to-head human evaluations. It is now available via the Yutori API. The following are excerpts from Yutori’s blog post on the release.
31 years ago, the modern web era began with the release of Netscape Navigator.
Last week, our team introduced Yutori Navigator, a state-of-the-art web agent that autonomously navigates websites on its own cloud browser to complete everyday tasks for you.
Navigator handles a wide range of web tasks, checking availability, comparing prices, filling out forms, making reservations, ordering food, and completing purchases, with pareto-domination over previous models on accuracy, latency, and cost.
Navigator is powered by Yutori n1, a pixels-to-actions LLM initialized from Qwen3-VL and trained via mid-training, supervised fine-tuning, and reinforcement learning (RL). During RL, n1 is not only trained on simulated web environments, but also on direct interactions with live websites. This uniquely allows n1 to learn the dynamics of real web environments and provides a scalable path for continuous improvement in the real world.
Navigator sets a new state of the art on browser-use benchmarks: 78.7% success rate on Online-Mind2Web3 and 83.4% on Navi-Bench, a new benchmark we are introducing. It also demonstrates strong real-world performance in Scouts, our consumer-facing web monitoring product, outperforming other computer-use models.
In addition to being the most effective, Navigator is also the most efficient agent, with per step latency that is 3.3x, 2.7x, and 2.0x faster than Claude 4.5, Gemini 2.55, and Claude 4.0 respectively.
We believe the internet can be an engine of productivity thanks to machines that pay attention and take action on your behalf.
See the company’s blog to learn more about the Navigator release. Navigator is now available via Yutori’s API. Sign up here to start using it.
AI News This Week
-
‘Sovereign AI’ Takes Off as Countries Seek to Avoid Overreliance on Superpowers (WSJ)
Smaller countries like South Korea are aiming to be self-sufficient by developing their own AI capabilities. Whether by acquiring advanced GPUs for localized computing infrastructure, developing large language models, or retaining engineering talent, these countries view autonomy in AI as a “life-or-death” priority.
-
Trump’s AI ‘Genesis Mission’: What are the Risks and Opportunities? (Nature)
The White House has launched the Genesis Mission, a national initiative to build AI models using the government’s computing resources and datasets from the 17 US national laboratories. Drawing a comparison with the Manhattan Project, the platform will focus on research into scientific challenges such as nuclear fusion, quantum science, and materials science.
-
Project Iceberg: Can AI Work with You? (MIT)
MIT research estimates that AI can automate 11.7% of work conducted in the U.S labour market. These findings were generated from Project Iceberg, an effort to create a digital twin of the U.S labour market by simulating 151 million workers individually. Researchers emphasize that the model does not predict specific layoffs, but rather reflects the percentage of wage value AI can automate. Separate research by MIT Sloan shows that AI adoption in enterprises correlates with faster revenue and employment growth.
-
New AI Model Enhances Diagnosis of Rare Diseases (FT)
Scientists have developed an AI model to assess the likelihood that a genetic mutation in a patient’s genome causes rare diseases. popEVE reflects new insights from researchers into how gene mutations affect protein synthesis across different species. The model reflects AI’s potential to deliver individualized care by leveraging genomic sequencing to analyze a person’s unique DNA.
-
Research: Super-Scalable Distributed Data Engine for Generalizable Computer Agents (MIT/UIUC/CMU/USC/UVA/UC Berkeley)
Researchers have released OSGym, software that can produce over a thousand copies of operating systems (OS) on which AI agents can be deployed simultaneously. OSGym sets up the necessary conditions for an agent to interact with the OS (via keyboard inputs, clicks, mouse movements, API-tool interactions) and evaluate its performance. As AI systems are deployed beyond chat interfaces, platforms like OSGym will make it easier to develop agents that use computers as humans do across different programs.
Radical Reads is edited by Ebin Tomy (Analyst, Radical Ventures)
