OpenAI's GPT-6 Astra Model Completes Portal Autonomously in 24 Hours for $571
OpenAI's GPT-6 Astra model has successfully completed the popular puzzle game Portal from start to finish without human intervention, a feat that took approximately 24 hours of autonomous operation. The total compute cost came to roughly $571 in tokens, showcasing the model's ability to plan, execute multi-step strategies, and adapt to in-game challenges over an extended period.
The accomplishment reflects significant advances in large language model reasoning and real-time problem-solving. Rather than relying on pre-scripted solutions, the model generated and refined its own approach to each puzzle chamber, drawing on its training to simulate logical planning typically associated with human players.
Researchers and observers note that completing a complex 3D puzzle game autonomously represents a meaningful benchmark for AI agents. It tests a model's capacity to maintain context across many interactions, handle spatial reasoning, and respond to feedback loops—all areas where earlier models struggled. The relatively low cost of the run also suggests that such tasks could become increasingly feasible as inference hardware improves and token prices decline.
While Portal is a game and not a real-world system, researchers view this type of test as a proxy for broader autonomous agent performance. The result adds to a growing body of evidence that frontier language models are becoming more capable at sustaining goal-directed behavior over hours rather than minutes.