GPT-6 Astra Beats Valve's Portal Solo After 24 Hours of Autonomous Play
An enthusiast ran an experiment where OpenAI's GPT-6 Astra played through the entire Portal game without human intervention, using a paused-game interface and screenshots to plan moves. The model made 3,336 tool calls and racked up $571.18 in API costs, though the user covered this with a $200 Codex Pro subscription. The feat highlights progress in multimodal AI, though the creator cautions it's not a formal benchmark.
The experimenter, CozyBlaze, orchestrated the run by keeping the game frozen while the AI processed visual data and player coordinates. Once the model formulated a sequence, the game resumed to execute the commands, resulting in 3,336 tool calls and $571.18 in token expenses, which were offset by a $200 Codex Pro plan.
CozyBlaze clarified that this accomplishment is not an official benchmark but echoes OpenAI's 2016 objective of a single agent conquering diverse games. The complete autonomous session spanned roughly 24 hours of streaming, condensed into a two-hour highlight reel, with the agent's code publicly accessible on GitHub for replication.
This demonstration could indicate that multimodal AI is advancing toward handling complex, sequential visual reasoning in dynamic environments. If such agents can navigate 3D spaces and solve logic puzzles autonomously, they may eventually aid in robotics, software testing, or virtual simulation. However, the substantial token costs and reliance on a paused interface suggest practical deployment remains economically constrained. Society could see accelerated automation in technical problem-solving, though real-world applicability outside controlled digital settings may still be limited.