GPT-6 Astra Completed Portal Solo, Experiment Cost $571

OpenAI's GPT-6 Astra model has independently completed the full, three-dimensional version of Valve's Portal, without any human help, hints, or a walkthrough. The experiment was carried out by an independent AI enthusiast known online as cozyblaze, and the result quickly spread across tech media as further proof of how fast the agentic capabilities of the latest language models are advancing.

The experiment's author connected the model to the game using a modified version of the SourcePauseTool alongside the Model Context Protocol, which handled the actual key presses and character control. A key part of the setup was pausing the game while the model analyzed the on-screen image and planned its next move, with play resuming only once Astra had decided on its next sequence of actions.

How the test worked

Astra did not use any simplified, text-based version of the gameplay. The model received only screenshots from the full 3D game and had to use them to read the space, plan portal trajectories, and solve multi-step physics puzzles, exactly as a human holding a gamepad would.

Because a language model's reasoning pace is far slower than the real-time physics of the game, the experiment's creator chose to pause the game engine while Astra was thinking. This let the model thoroughly analyze the layout of a room and plan a sequence of portal jumps before the action continued.

The model didn't finish the game in two hours as the recording suggests, the game was paused every time GPT-6 Astra needed to think, and those pauses were cut from the video

How much it cost

Completing the full game took the model about 3,336 tool calls, and the total cost of tokens consumed via the API came to $571.18. The test's author, however, was working from a $200 monthly Codex Pro subscription, so the real cost from his perspective was far lower than the calculated market value of the computing resources used.

The high cost of a single playthrough shows that despite the growing capabilities of agentic models, using them in practice for tasks requiring long, continuous real-time reasoning remains expensive. Critics commenting on the result noted that the writeups describing the experiment lacked full technical details of the setup, making independent verification of the test difficult.

Context for the gaming industry

Portal, released by Valve in 2007, has long been considered one of the most demanding computer games in terms of spatial reasoning, requiring a player to simultaneously manage level geometry, movement physics, and puzzle logic built around moving objects between portals. The experiment was compared to earlier, far less successful attempts by AI models at video games, about a year and a half earlier the Claude model struggled badly to complete the much simpler Pokemon.

The progress shown by Astra fits into a string of reports about the growing capabilities of this particular model beyond typical office applications, OpenAI had previously reported, among other things, that Astra solved ten open mathematical problems. The ability to navigate a complex 3D environment on its own, plan multi-step sequences, and correct errors without human intervention is seen by researchers as a significant indicator of language models' general spatial reasoning abilities, extending beyond pure text processing.

For game developers and platforms like Steam, the experiment's result is another signal that AI models are becoming increasingly capable of handling tasks that require visual perception and real-time planning, not just text or code analysis. At the same time, the high cost and time needed to complete a single game show that widespread, cheap use of such agentic capabilities in entertainment is still a long way off.