GPT-6 Astra Beats Portal Without Human Help

An autonomous AI agent powered by GPT-6 Astra has successfully played through the entire video game Portal without any human assistance. Developed by a creator known as cozyblaze, the agent navigated the complex physics-based puzzle game from the opening chamber to the end credits in 23 hours and 43 minutes. The feat represents a significant step toward OpenAI's 2016 vision of creating a single, generalized agent capable of mastering diverse gaming environments.

To achieve this, the developer integrated GPT-6 Astra with the game using the Model Context Protocol (MCP) alongside a modified version of SourcePauseTool. This setup paused the gameplay whenever the model needed to process information. During these pauses, the agent analyzed real-time screenshots, the player's coordinate position, and the camera angle. After evaluating the visual and spatial data, the model selected its next inputs and resumed the game to execute the actions.

Running the model through the entire game required substantial computational resources. At GPT-6 Astra's standard list price, the total token consumption would have cost at least $570. However, cozyblaze managed to complete the run using a $200 Codex subscription. The developer has made the underlying code and documentation publicly available on GitHub for other researchers to study and replicate.

For AI practitioners, this milestone demonstrates the growing viability of multi-modal models in solving complex, spatial-reasoning tasks in real-time environments. While the reliance on pausing the game highlights current latency limitations, it proves that modern LLMs can translate visual and coordinate data into precise physical controls. As cozyblaze remarked on X, GPT-6 Astra represents "the worst model we'll ever get," suggesting that future iterations will only become faster, cheaper, and more capable of handling dynamic, unscripted environments.