Korean Startup’s FluidGPT Achieves First Perfect Score on AndroidWorld

FluidGPT, a Korean startup founded by a KAIST professor, has developed the first mobile AI agent to achieve perfect accuracy on AndroidWorld, a benchmark created by Google DeepMind‘s research team.

FluidGPT completed all 137 tasks in the evaluation with 100% accuracy, placing it ahead of agents developed by Google’s Artemis research team, Microsoft Research, Alibaba, and other major research groups.

Testing Mobile AI Capabilities

AndroidWorld tests mobile AI agents on real-world tasks including app navigation, screen recognition, contextual understanding, and multi-step decision-making. The benchmark goes beyond button-pressing to evaluate whether agents can manage complex sequences across multiple applications while maintaining context—capabilities considered essential for autonomous smartphone operation.

Mobile environments present particular challenges for AI systems due to varied screen layouts, frequent app transitions, and complex user interfaces. Industry observers regard strong AndroidWorld performance as an indicator of whether agents can move beyond demonstration-level capabilities to complete practical tasks.

Leadership Perspective

Shin In-sik, CEO of Fluiz and a professor at KAIST‘s School of Computing, said the perfect benchmark score demonstrates that FluidGPT possesses competitive accuracy and practical capability in the global mobile AI agent market. He said the company aims to become a leading player in on-device AI and the mobile operating system space.