GPT-6 Astra finished Portal on its own - 3,336 moves, nearly a full day and a $571 bill

Eighteen months ago Claude tried to beat Pokémon Red on the ClaudePlaysPokemon channel. It earned three badges and then spent over eighteen hours stuck in Mt. Moon, walking in circles and losing the progress it had made. Watching it was an exercise in patience rather than a display of capability. Last Monday GPT-6 Astra finished the whole of Portal.

In the recording posted by the user cozyblaze on X, the character moves like an aimbot - the camera snaps straight to the exact spot where the portal has to go, with no correction. In a sense that is true, because the model steers the camera with numbers, not a mouse. Except those two hours of video are an edit. It really took almost a full day and 3,336 tool calls, and the game was frozen every time the model stopped to think. The pauses were cut out.

The setup was simple. The model received screenshots and player coordinates, a modified SourcePauseTool froze the game while it deliberated, and a Model Context Protocol server pressed the keys. Astra planned a sequence of moves, the game resumed, executed them and stopped again. That went on through the entire campaign, chamber after chamber.

The cost was $571.18 in tokens - with Astra priced at $10 per million input tokens and $50 per million output, a full day of continuous work adds up quickly. The author clarified, though, that he did not pay that separately: it all fit inside his $200-a-month Codex Pro subscription. As a way to finish a game from 2007 it is still a poor deal.

It is worth being clear about what was actually measured. Every chamber in Portal has had its solution written up online for well over a decade, so this is not a test on unfamiliar puzzles - the model may well have known what to do before it saw the first screen. The real difference from the Pokémon run lies elsewhere: the agent held on to a single goal across 3,336 consecutive decisions and did not fall apart along the way. That is the ability that earns money in office work, not in games.

Tool call - a single action the model asks an outside program to perform, for example "press W for half a second" or "give me a screenshot". Model Context Protocol - an open standard for connecting AI models to external programs and data; here it acted as the go-between for the model and the keyboard.