OpenAI's GPT-6 Astra Completes Portal Autonomously in Costly 24-Hour Run
An experiment by user cozyblaze demonstrates the spatial reasoning of OpenAI's latest model, though high API costs and slow processing remain significant hurdles.
OpenAI's GPT-6 Astra has successfully completed the puzzle game Portal without human intervention, marking a significant milestone in autonomous AI agency. The experiment, conducted by a user named cozyblaze, proves that large language models can now navigate and solve complex spatial puzzles in a dynamic 3D environment.
To achieve the victory, the AI utilized a combination of screenshots, character coordinates, and control tools to analyze the game state and execute movements. According to reports from VGTimes and iXBT, the process required 3,336 individual tool calls and requests. The run was neither fast nor cheap: the AI took nearly 24 hours of real-time processing to finish the game, incurring a total API cost of $571.18.
The Power of Astra
Released in September 2026, GPT-6 Astra is OpenAI's flagship model specifically engineered for high-demand, end-to-end computer use. A critical component of its success in this experiment is its massive 1,050,000 token context window, which allows the model to maintain a vast amount of information about its previous actions and the environment's layout. This capability is essential for games like Portal, where solving a puzzle often requires remembering the position of an object or a portal entrance from several rooms prior.
The Cost of Autonomy
While the technical achievement is impressive, the economics of the run highlight a stark divide between capability and practicality. The $571.18 expenditure for a single playthrough far exceeds the retail price of the game itself. Furthermore, the 24-hour timeframe illustrates that while the AI can eventually solve these problems, it does so at a pace that is currently impractical for real-time application or consumer-grade automation.
The Path Forward
This demonstration underscores the evolving role of LLMs as autonomous agents capable of complex problem-solving in visual environments. The ability to translate visual data into precise tool calls suggests a future where AI can manage software interfaces and physical simulations with minimal oversight. However, the industry must now address the efficiency gap. Until API costs drop and processing speeds increase, the vision of a seamless, autonomous AI agent will remain a costly luxury rather than a widespread utility.