The run
A developer publishing as cozyblaze has released the code and a sanitised log of an agent that played Valve’s 2007 puzzle game Portal from beginning to end using OpenAI’s GPT-6 Astra, without human intervention after the initial objective.
The run summary in the repository dates the attempt to 4–5 September and puts the duration at roughly 23.7 hours. It records 433,210,793 input tokens, of which 426,415,104 were cache reads, alongside 1,598,341 output tokens including 1,181,102 of reasoning output. The agent issued 3,210 execution commands and took 40 screenshots.
How it worked
The setup is a pause-and-step loop rather than a model playing in real time. A modified SourcePauseTool halts the game; the agent receives a screenshot along with the player’s position and camera angle; it decides its inputs; the game advances a set number of ticks; the cycle repeats. Commands reached the game through a local MCP server.
That design matters for what the result means. Nothing here demonstrates reaction speed or continuous control. What it demonstrates is a model holding a goal across tens of thousands of steps, reading a 3D scene from still frames well enough to navigate it, and recovering when a plan fails.
The caveats, stated by the author
The repository is candid about its limits. The run hit capacity interruptions and had to be resumed after errors. The model was run at maximum reasoning effort and later switched to a faster mode. Personal information and embedded screenshots were stripped from the published log, and the game files are not included.
This is one developer’s demonstration on one game, not a benchmark, not a controlled evaluation, and not something OpenAI has published or verified. No independent party has reproduced it. The evidence is a log the author chose to publish — which is more than most such claims come with, and still not the same as a third-party result.
Why anyone should care about a nineteen-year-old game
Portal is a useful test precisely because it is old. Its puzzles require the player to reason about physics and space — where a portal must be placed so that momentum carries you somewhere you cannot walk — and they cannot be brute-forced by reflexes. Games as an AI proving ground go back a long way; OpenAI set itself the goal of a single agent solving many games a decade ago.
The cost figure is the part with commercial relevance. Nearly all of those 433 million input tokens were cache reads rather than fresh input, which is what makes a run of this length affordable at all. The Decoder estimated the cost at around $570 at standard pricing; the repository does not state a figure, and that estimate is theirs, not the author’s.
What to watch
Whether anyone reproduces the run, and whether the same scaffolding transfers to a game the model has not seen described anywhere in its training data. Portal has been written about extensively for nineteen years, including complete walkthroughs. Ruling out recall rather than reasoning would need a puzzle nobody has published a solution to.
