GPT-6 Astra completes Portal game without human assistance

Summary

OpenAI's GPT-6 Astra model completed the puzzle-based FPS game Portal without human intervention. The process took 23 hours and 43 minutes, involved 3,336 tool calls, and incurred an API cost of $571.18. The model utilized a custom MCP-based infrastructure that paused the game during its reasoning phase.

Why it matters

Completing a 3D puzzle game featuring spatial reasoning, timing-sensitive mechanics, and a multi-stage story boss (GLaDOS) pushes large language models (LLMs) far beyond text benchmarks. This demonstrates the model's ability to engage in concrete, sequential decision-making processes. However, the $571 token cost and 23-hour runtime highlight that this remains more of a research curiosity than a practical capability. Furthermore, the model's reliance on custom-built pause and resume infrastructure, such as SourcePauseTool, proves it cannot operate in a real-time, unsupported environment.

Highlights

• 3,336 tool calls and repeated attempts indicate that the model solved the Portal game through brute-force trial-and-error rather than elegant planning.
• The SourcePauseTool plugin freezes the game during inference, ensuring the model is never exposed to real-time pressure or dynamic state changes.
• The full setup, cleaned session logs, and a three-part video recording (VOD) are publicly available on GitHub and YouTube for independent verification.

Critical take

The claim of 'no human help' is a generous description. Human engineers designed the MCP controller, the pause/resume patch, and the entire underlying infrastructure. The model's 23-hour, $571-long process of repeated retries resembles a random search. This is vastly different from the fluid, insight-driven gameplay a human can achieve in under an hour. While it may be fair to call this a milestone, the cost-to-capability ratio and the required scaffolding infrastructure suggest it is a proof of concept rather than evidence of general game intelligence.