Notebookcheck Logo

GPT-6 Astra AI model beats Valve's Portal game solo in 24 hours

GPT-6 Astra AI model finishes Valve's Portal on its own
ⓘ cozyblaze/X
GPT-6 Astra AI model finishes Valve's Portal on its own
There have been a number of "AI plays old game" stories at the dawn of the AI era, and they typically end with a chess engine losing spectacularly to a random Atari opponent. The new AI models like OpenAI's GPT-6 Astra, however, are built differently and can beat Valve at its own game.

An AI enthusiast going by the handle CozyBlaze recently set OpenAI's new flagship model, GPT-6 Astra, loose on Valve's Portal. The AI model ran the full 3D puzzle game, not the stripped text version, without a walkthrough or scripted assist. The challenge included just the model, a modified SourcePauseTool, and the Model Context Protocol handling the actual button presses.

Whenever Astra needed to think, however, the game froze and continued once the agentic AI had worked out its next sequence of moves to play out in real time. That's a rather clever workaround for the gap between an LLM's reasoning speed and a game engine's framing physics. That is why the finished video, trimmed down to about two hours for demonstration purposes, actually represents roughly 24 hours of raw uncompressed gameplay.

GPT-6 Astra Portal gameplay costs

Astra finished the whole game in about 3,300 tool calls, at an API cost of less than $600, though CozyBlaze notes the run was actually covered by a $200 a month Codex Pro subscription.

CozyBlaze himself plays down the achievement, eschewing the temptation to call it a benchmark as so many others that put AI through unorthodox paces. Watching a general-purpose AI agent explore an unfamiliar 3D environment and reason its way through spatial puzzles end to end, however, feels like a real step toward the kind of universal gameplay agent that OpenAI has been teasing for a decade or so.

As for the competition, Anthropic's Claude models have been focusing more on computer use and agentic coding tasks rather than game environments specifically, while Google's Gemini agents are leaning into multimodal tool use across browsers and apps. China's DeepSeek agent has mastered efficient reasoning and coding benchmarks, rather than open tasks like navigating a 3D game world, so for now OpenAI's latest GPT-6 Astra AI model is in a league of its own when it comes to spatial reasoning inside a physics-driven environment like Valve's Portal puzzle. 

Source(s)

Google LogoAdd as a preferred source on Google
Mail Logo
static version load dynamic
Loading Comments
> Expert reviews and news on laptops, smartphones and tech innovations > News > News Archive > Newsarchive 2026 09 > GPT-6 Astra AI model beats Valve's Portal game solo in 24 hours
Daniel Zlatev, 2026-09- 7 (Update: 2026-09- 7)