OpenAI's GPT-6 Astra Autonomously Beats 3D Puzzle Game Portal
An AI enthusiast named CozyBlaze successfully enabled OpenAI's GPT-6 Astra model to autonomously complete Valve's 3D puzzle game Portal. The AI agent completed the game over 24 hours of real-time play using the Model Context Protocol (MCP) alongside a modified game-pausing tool. This achievement demonstrates significant progress in multimodal AI agents performing complex 3D spatial reasoning, dynamic environment interaction, and long-horizon planning. It advances OpenAI's long-held vision of developing general-purpose agents capable of tackling diverse game environments and real-world execution tasks. The run required 3,336 tool calls and accumulated API costs of $571.18, which were covered under a $200/month Codex Pro subscription. The setup used SourcePauseTool (SPT) to freeze the game while Astra analyzed screenshots and coordinate data, unpausing only to execute planned sequence actions.
## BACKGROUND
Model Context Protocol (MCP) is an open-source standard created to unify how large language models interact with external tools, applications, and contextual data sources. Portal is a physics-based 3D puzzle game by Valve requiring players to solve spatial puzzles using momentum and teleportation portals. SourcePauseTool (SPT) is a plugin used by speedrunners and developers to programmatically control game pausing and input execution in Source engine games.