Edge AI Daily Morning Report (September 8th)

On September 8, 2026, Edge AI Daily reported that the GPT-6 Astra model achieved autonomous completion in the game “Portal”. The experiment was conducted by enthusiast CozyBlaze using the MCP protocol to connect with SourcePauseTool. It took 24 hours, involved 3,336 tool calls, and incurred $571.18 in API fees. This experiment marks a shift from AI’s ability to “understand the world” to its ability to “operate the world”. In the OSWorld 2.0 benchmark test, GPT-6 Astra achieved a completion rate of 72.6%, a 47% improvement over previous versions; ScreenSpot-Pro’s accuracy in locating screen elements reached 92.7%. This experiment extends reasoning capabilities to three-dimensional physical space, providing valuable reference for robotics companies like Figure AI and facilitating the implementation of digital employees and automated software testing.