AI Navigates Complex Interactive Virtual Environments

AI Navigates Complex Interactive Virtual Environments

Tom’s Hardware ⊳ An enthusiast known as CozyBlaze has demonstrated OpenAI’s latest GPT-6 Astra model completing Valve’s puzzle game Portal entirely autonomously, using Model Context Protocol (MCP) and a modified SourcePauseTool to control inputs. GPT-6 Astra completed Portal in around 24 hours, navigating 3D spaces and solving puzzles using screenshots, positional data and tool-based controls. The experiment required 3,336 tool calls and generated an API cost of $571.18. Astra paused the game during reasoning before executing planned movements, demonstrating how a general-purpose AI agent can sustain planning and interaction across a complex virtual environment. Its creator cautioned against treating the achievement as a formal AI benchmark.
⊲ CozyBlaze’s use of Model Context Protocol to steer GPT-6 Astra through Portal hints at MCP becoming a standard bridge between language models and simulated environments. ⊳
⊲ Image – Portal Game/Valve⊳
www.simulation.news