GPT-6 Astra finished Portal alone, and the game was paused

A developer who posts as cozyblaze set GPT-6 Astra a single goal and let it play Portal. The model reached the credits without further human input, and The Decoder reports the run took about 23 hours and 43 minutes.

The game was paused for most of that time.

“GPT-6 Astra is the worst model we'll ever get.”

— cozyblaze, developer, The Decoder, 8 September 2026

Quote source: cozyblaze, developer, via The Decoder

The game was paused while it thought

Astra drives the game through the Model Context Protocol and a modified version of SourcePauseTool. The loop runs four steps:

1. The tool pauses Portal.
2. The model receives a screenshot, the player position and the camera angle.
3. It picks its inputs.
4. The tool lets the game run again.

So the model's clock ran for nearly a day while the game's clock barely moved. The published video cuts the pauses out, which is what makes the footage look like play rather than correspondence.

That is the honest frame. Astra did not react to a real-time game. It played a turn-based version of one that it created for itself, and the achievement sits in planning across thousands of turns without a person stepping in.

The loop also means the model never sees motion. It gets stills, a coordinate and an angle, then commits to inputs it cannot watch execute. Portal is a physics puzzle built around momentum you carry through a portal, so solving it from photographs is a harder problem than solving it with your eyes on the screen.

What the GPT-6 Astra Portal run costs to repeat

Token usage came to at least $570 at Astra's list price. Cozyblaze ran it on a $200 Codex subscription, so the experiment cost the model's operator more than it cost the person doing it.

Nobody sells Portal completions, so the number matters only as a unit. Twenty-four hours of continuous agent work on a task with a clear win condition and free retries costs a few hundred dollars in tokens. Most useful work has neither the clear win condition nor the free retries.

A goal set in 2016

OpenAI set itself the goal of solving many different games with one agent back in 2016, when it released the Universe platform. Ten years later a hobbyist did it for one game on a consumer subscription, using a general model that was never trained for it.

The code and documentation sit on GitHub, so the run is reproducible by anyone willing to spend the tokens. That matters more than the result. A demonstration you cannot repeat is a claim, and this one is not.

We covered Astra's safety overview last week, where OpenAI reported perfect scores on exploitation benchmarks, and the apology for its rollout the day after. A model that can be handed a goal and left alone for a day is the same capability read two ways.

This article is for informational purposes only and does not constitute investment advice.