Astra, Opus 5.5, and other Frontier Models Demonstrate Jagged Performance Across SoTA Agentic Tasks from Web Browsing to Robotics

A team of researchers led by Yangyue Wang, with co-authors Harshvardhan Sikka, Pranav Guruprasad and Sudipta Chowdhury of Fig and the Georgia Institute of Technology, reported that frontier models including Astra and Opus 5.5 show jagged performance across state-of-the-art agentic tasks ranging from web browsing to robotics. The work evaluates these models on agentic benchmarks and finds uneven capability across task types.

Yangyue Wang 1 , Harshvardhan Sikka 1, 2 , Pranav Guruprasad 1 , Sudipta Chowdhury 1 1 Fig ; 2 Georgia Institute of Technology .