An agent running on GPT-6 Astra played Portal from start to finish with no human help, in under 24 hours. The mechanism is a pause loop: the game stops, the agent receives screenshots, reads the character's position and the camera angle, selects its actions, and play resumes. Every one of those pauses was cut out of the recording. The code and documentation are on GitHub, published by the developer, Cozyblaze.
Source: the-decoder.com
Cozyblaze pointed back to 2016, when OpenAI set itself the goal of training a single agent capable of solving a wide range of games. Problems remain, he said, but finishing an entire game unaided shows a small piece of that idea working. His summary of the model was the industry's standard formula: GPT-6 Astra is "the worst model we will ever get."
The pauses are the whole result, and they deserve more attention than the completion time. Portal is a real-time first-person game whose difficulty is distributed across two axes — spatial reasoning about where a portal has to go, and execution under a physics clock. Freezing the game at every decision removes the second axis entirely. What this run demonstrates is that a model can look at a static frame, work out where it is standing and what the geometry requires, and choose the next move correctly enough, hundreds of times in sequence, without drifting into an unrecoverable state. That is genuinely hard and was not reliably possible a short time ago. It is also not playing Portal in the sense most people will read into the headline.
Cutting the pauses from the video is an editing decision, not a deception — nobody wants to watch dead frames — but it means the footage shows a capability the system does not have. The honest framing is that an agent solved Portal as a turn-based puzzle game with a very large number of turns.
The question the release leaves open is how many runs it took. With the code public, the interesting number is not 24 hours but attempts: whether this was the first complete playthrough or the one that survived after a stack of failures, how often the agent stalled, and how much inference the whole thing cost. A single successful trajectory tells you a ceiling has been touched. A success rate tells you whether the capability is real. The former was announced; the latter was not.
"The worst model we will ever get" is a claim about every future model dressed as modesty about the current one, and on this evidence the useful benchmark for the next one is already obvious. It is not a longer game or a faster finish. It is whether the pauses can come out of the loop instead of out of the recording.