OpenAI’s GPT-6 Astra has achieved a notable milestone in autonomous agent capability by completing the entire puzzle-platformer Portal without human intervention. Developer cozyblaze shared the results on X, detailing a run that lasted approximately 23 hours and 43 minutes from start to credits. This wasn’t a pre-scripted speedrun but a genuine attempt by the model to interpret visual data, reason through spatial puzzles, and execute inputs in real time. The feat marks a significant step toward OpenAI’s 2016 goal of creating a single agent capable of solving diverse games, though it comes with substantial computational costs.
The technical setup relied on the Model Context Protocol (MCP) and a modified version of SourcePauseTool. This tool allowed the agent to pause the game engine while it processed information. During these pauses, Astra analyzed screenshots, player coordinates, and camera angles before deciding on its next move. Once the decision was made, the game resumed. While the final video edit removes these thinking pauses to show continuous gameplay, the underlying process highlights the current latency and resource intensity of such autonomous agents. It is a slow, deliberate form of intelligence rather than the instant reflexes of a human gamer.
Financially, the experiment was expensive. At Astra’s list price, the token usage for this single playthrough would cost at least $570. Cozyblaze managed to keep costs down to a $200 Codex subscription, but the baseline expense underscores why widespread autonomous gaming agents aren’t yet viable for consumers. Despite the cost, the developer offered a provocative perspective on the model’s performance, calling GPT-6 Astra “the worst model we’ll ever get.” This suggests that future iterations will be significantly more efficient and capable, making this clumsy 24-hour grind look like a primitive ancestor of what’s coming next.
While problems remain in handling complex, dynamic environments, this demonstration proves that large language models can now maintain long-horizon goals across thousands of discrete actions. It moves beyond simple chat or code generation into persistent, interactive world navigation. The industry should watch closely as these agents evolve from costly novelties into practical tools for software testing and virtual interaction. If Astra is indeed the baseline, the ceiling for autonomous digital agents is rising faster than most expected.
