User X under the nickname Miłosz Szewsky asked ChatGPT-6 Astra to complete a hardcore platformer Cuphead. After several hours and $700 worth of tokens, the result was, to put it mildly, not the most impressive.
Judging by the short video, the AI could not even master the training.
But the new model did a much better job with Portal. User cozyblaze connected Astra to a Valve puzzle game and tasked her with completing the entire game on her own. The model received screenshots, character coordinates, and the ability to control the game through connected tools.
As a result, she actually made it to the finals – although it took almost 24 hours, including pauses for reasoning.
https://www.youtube.com/watch?v=g5u2y0BwRJ0
AI GPT-6 collects other people’s games on the cheap, but worse and without a future
During the passage of Portal, Astra made 3336 tool calls, and the API cost was $571.18. At the same time, the author of the experiment himself notes an important nuance: Portal was released many years ago and was analyzed in such detail on the Internet that the model could already have a significant amount of information about the game.
