GPT-6 Astra Autonomously Completes Portal Game in 24 Hours, Costs $571 in API Calls

GPT-6 Astra Achieves Autonomous Portal Completion

OpenAI's GPT-6 Astra has successfully completed Valve's Portal, a complex 3D puzzle game, entirely on its own in approximately 24 hours1

1

2

2

. AI enthusiast CozyBlaze conducted the experiment, demonstrating what many consider a significant leap in multimodal AI capabilities and general-purpose agents. The large language model had to map 3D game spaces, understand complex puzzles, and plan solutions before executing them, showcasing advanced spatial reasoning and problem-solving abilities that extend beyond text-based tasks1

1

.

Technical Implementation and Token Usage

The model controlled Portal through the Model Context Protocol combined with a modified SourcePauseTool, a setup that allowed the game to pause while GPT-6 Astra processed information1

1

. During these thinking periods, the AI received screenshots and player character position data. Once it formulated a plan, the game resumed and Astra executed its planned moves with precision that appeared aimbot-like in speed and accuracy2

2

. The entire playthrough required 3,336 tool calls and generated an API cost of $571.18 in tokens1

1

. CozyBlaze later clarified these costs were covered under their $200 Codex Pro subscription1

1

. While the edited highlight video runs approximately 2 hours, the full gameplay with all thinking pauses removed totaled around 24 hours of processing time1

1

.

Progress in AI Gameplay Capabilities

This achievement marks substantial progress compared to earlier AI gaming attempts. About a year and a half ago, Claude struggled with basic tasks in Pokémon, frequently getting stuck and undoing progress on the Twitch channel ClaudePlaysPokemon2

2

. The contrast highlights how rapidly LLM capabilities have evolved. CozyBlaze acknowledged that while problems remain and this shouldn't be viewed as a definitive AI benchmark, watching a general-purpose agent autonomously navigate through an entire game represents a glimpse of OpenAI's original 2016 vision becoming reality—when the company mentioned an eventual goal of creating a single agent capable of solving a variety of games1

1

. The Portal Agent code is now available on GitHub for others to explore and experiment with1

1

.

OpenAI's State-of-the-Art Model

GPT-6 Astra became OpenAI's flagship model earlier this month, with the company claiming it represents a new generation of intelligence1

1

. OpenAI positions the model as state-of-the-art across multiple domains including computer use, browsing, software engineering, cybersecurity, science, and professional work1

1

. The company hopes GPT-6 Astra will bring the world of large language models closer to the AGI era2

2

. This Portal completion demonstrates the model's ability to handle complex spatial reasoning and sequential decision-making in real-time environments, capabilities that extend far beyond traditional text processing. Watch for more demonstrations of Astra's capabilities in software engineering and computer use tasks as developers continue testing its limits across different domains.