An indie developer wired OpenAI's new flagship to Valve's 2007 puzzler Portal via a paused game loop, with a $200 subscription absorbing the $571 in API tokens and the indie developer's caveat that it isn't a benchmark.
An independent AI enthusiast named CozyBlaze wired OpenAI's newly released GPT-6 Astra to Valve's 2007 first-person puzzle game Portal and let the model play through the whole campaign on its own. The run, published to GitHub and streamed live, finished in roughly 24 hours across about two hours of edited highlights.
The game stayed paused while Astra read screenshots and the player character's position, then unpaused to execute a planned input. That pause-think-execute loop runs through the Model Context Protocol (MCP) and a modified version of SourcePauseTool, repeated 3,336 times. The model itself launched earlier this month as OpenAI's flagship.
The headline cost is $571.18 in API tokens, the figure Tom's Hardware and The Verge picked up. CozyBlaze clarified in the same thread that the figure is the API list price, not cash outlay. The run was absorbed by a $200-per-month Codex Pro subscription, so the marginal cost to the runner was effectively zero.
CozyBlaze frames the demo as "a small glimpse" of an OpenAI goal first articulated in 2016, and is explicit the run "shouldn't be viewed as an AI benchmark." One game, one controlled harness, one subscription, one open repo: not a general-agent milestone.