Particle.news

OpenAI’s GPT‑6 Astra Autonomously Completes Valve’s Portal Using Pause‑and‑Execute Tools

The experimenter says the model controlled the game through a pause-and-execute pipeline to demonstrate agent‑style multimodal control and presents the run as an informal single demonstration.

Overview

  • An AI enthusiast reported that GPT‑6 Astra autonomously completed Valve’s 3D puzzle game Portal by issuing 3,336 tool calls through a tooling pipeline that paused the game while the model 'thought' and then executed actions.
  • The full playthrough was recorded across roughly 24 hours of live footage with edited highlights of about two hours, which reflects long pauses while the model received screenshots and planned moves.
  • Estimated API charges for the run totaled about $571, though the experimenter said those costs were covered by an existing subscription service rather than paid per run.
  • Technically, the setup used the Model Context Protocol (MCP) and a modified SourcePauseTool to deliver game state to the model, pause execution while it planned, and then apply its chosen inputs.
  • Observers and the experimenter stress this is a single, informal demonstration that shows progress toward OpenAI’s single-agent vision but does not prove robustness, reproducibility, or readiness for general deployment.