GPT-6 has finished Portal, but the bill is steep

GPT-6 Astra, the latest model developed by OpenAI, has managed to complete entirely autonomously the famous Portal by Valve.

Although the feat demonstrates a very rapid maturation of these systems’ computational and action capabilities within complex three-dimensional environments, the milestone has come with a substantial cost.

To reach the end credits, the operation has indeed generated an expenditure of over $500 in API calls.

OpenAI’s latest model finishes Portal, here’s the video

Until about a year and a half ago, the interaction between LLMs and video games yielded radically different results. Many will remember the live streams on the Twitch channel ClaudePlaysPokemon, where the artificial intelligence struggled greatly to advance in a Pokémon game.

On that occasion, the system frequently froze, erased its own progress, and offered an experience that was at times frustrating for viewers.

Today the situation appears profoundly changed. OpenAI places enormous hopes in GPT-6 Astra, presenting it as the vehicle to push development toward AGI. Whether or not it is the definitive step toward AGI, the demonstration offered with Valve’s masterpiece confirms an impressive leap in puzzle-solving logic.

The news of the feat emerged thanks to the outlet VideoCardz, which intercepted a post published on the social network X by user cozyblazex. The shared video shows GPT-6 Astra easily solving Portal’s physics.

The model’s movements within the map look almost unnatural due to their cold perfection: the camera movements and interactions are extremely precise.

The published video states that the gaming session ended in only two hours, but this figure hides a tiny footnote. During the actual game, the action was regularly paused to allow the system to process the code and the next move; these computation times required by the API calls were removed during video editing.

A hefty bill for an exceptional result

Despite the undeniable success of the operation, the autonomous management of the video game by the model required a massive ongoing use of computational resources. At the end of the challenge, the user cozyblazex found themselves having to pay $571.18, entirely consumed by the necessary API calls.

Certainly not the most economical way to approach a gaming milestone, yet the experiment has enormous technological value.

This practical execution highlights the speed of learning of such architectures, leaving a strong curiosity about how linguistic and multimodal systems will interact with even more articulated gaming works within a couple of years.