GPT-6 Astra completed the full game of Portal without human assistance in 23 hours and 43 minutes. Developer Cozyblaze confirmed the run on X, noting the model reached the credits after an initial goal was set. The process involved OpenAI‘s MCP and a modified SourcePauseTool, which pauses gameplay to allow the agent to analyse screenshots and player positions before resuming. Token consumption for the list price totals at least $570, though Cozyblaze utilised a $200 Codex subscription to reduce costs. The video evidence removes the pauses to show continuous progress.
This achievement demonstrates that large language models can operate as autonomous agents for complex tasks previously thought to require significant human oversight. It validates a 2016 OpenAI objective to solve multiple games with a single agent, even though broader challenges persist. The result suggests future applications where software handles intricate decision-making loops independently.
* Run time was 23 hours and 43 minutes
* Token cost exceeds $570 at list price
* Cozyblaze called the model the worst they will ever get
