AI beat Portal for $571 — but it took 24 hours to do it
An AI model just finished Portal from start to finish without any human help — and the bill came to $571.18. GPT-6 Astra, developed by OpenAI, worked through every puzzle chamber in Valve's classic independently, making 3,336 API calls along the way. The catch: what looks like a two-hour gaming session actually took nearly 24 real-world hours.
The experiment was run by an enthusiast going by the handle cozyblaze on X. The setup gave Astra access to game screenshots, player coordinates, and action controls. The model processed each frame sequentially — pausing gameplay while it analyzed visuals and decided its next move, effectively turning a fast-paced first-person puzzler into something resembling a very slow turn-based strategy game.
What it actually proved
The result is genuinely impressive: Astra navigated 3D space, solved multi-step physics puzzles, and never got stuck on a chamber. That puts it well ahead of where AI agents were 18 months ago — when Anthropic's Claude reportedly failed to make meaningful progress in Pokémon on a dedicated Twitch channel.
And... GPT-6 Astra has autonomously completed Portal! I didn’t expect this to happen so soon, but I’m glad we've made so much progress here.
— cozyblaze (@cozyblazex) September 5, 2026
I was reminded that back in 2016, one of OpenAI’s technical goals was to “solve a wide variety of games using a single agent.” pic.twitter.com/2nVREdCbMI
There's an important caveat, though. Portal is one of the most documented games on the internet, meaning Astra almost certainly encountered its puzzles during training. That's a very different challenge from dropping an AI into an unfamiliar game it's never seen before. Completing a well-mapped puzzle box is not the same as general problem-solving in novel environments.
The price tag
The $571.18 cost is the number that stops this story from being a straight-up triumph. That's roughly three times the price of a decent gaming PC component upgrade, and around 23 times what Portal itself costs on Steam. The edited highlight video runs two hours — the actual inference time was close to a full day, with the game paused between each model decision.
For context, $571 in API fees to beat a game a human speedrunner clears in under nine minutes is a useful reminder that inference costs, not raw intelligence, remain the real bottleneck for autonomous AI agents.
The broader stakes are also worth noting. Anthropic is currently facing significant copyright claims from authors over training data — a precedent that raises questions about whether games publishers could or should restrict model access to their titles going forward.
