OpenAI's GPT-6 Astra autonomously completes Portal game in 24 hours for $571

Sep 6, 2026

Developer cozyblaze demonstrated OpenAI's new flagship model, GPT-6 Astra, autonomously completing Valve's Portal in 23 hours and 43 minutes. Using the Model Context Protocol and a modified SourcePauseTool to pause the game while thinking, the run generated 3,336 tool calls and cost $571.18 in API tokens, covered by a $200 Codex Pro subscription. Meanwhile, Nvidia CEO Jensen Huang declared that artificial general intelligence (AGI) has arrived, stating Astra was trained on over 100,000 Nvidia Grace Blackwell NVLink72 systems.

GPT-6 Astra Portal completion

▪The GPT-6 Astra model controlled the game Portal using the Model Context Protocol and a modified SourcePauseTool that paused gameplay while the model calculated moves.
▪Developer cozyblaze used OpenAI's GPT-6 Astra model to autonomously complete the video game Portal in 23 hours and 43 minutes without human intervention.
▪The unedited gameplay streams of the GPT-6 Astra Portal run totaled approximately 24 hours, which developer cozyblaze condensed into a two-hour edited video.
▪During thinking pauses in the Portal run, the GPT-6 Astra model analyzed screenshots, player positions, and camera angles before executing planned input sequences.

API token costs

▪Developer cozyblaze clarified that the API costs for the GPT-6 Astra Portal run were covered under their $200 Codex Pro subscription.
▪The total API token cost for OpenAI's GPT-6 Astra model to complete the game Portal was $571.18, which required 3,336 tool calls.

Jensen Huang AGI claims

▪Nvidia CEO Jensen Huang congratulated OpenAI on social media, highlighting the progression from ChatGPT to o1 to Astra within a four-year period.
▪Nvidia CEO Jensen Huang declared that artificial general intelligence has arrived, crediting the launch of OpenAI's GPT-6 Astra model.

Nvidia GPU infrastructure

▪Nvidia CEO Jensen Huang stated that GPT-6 Astra was trained on over 100,000 Nvidia Grace Blackwell NVLink72 systems, revising an initial claim of 300,000 systems.
▪Nvidia's NVLink72 architecture connects 72 chips together to function as a single unified system, representing a shift from shipping eight processors per board.
▪A $100 billion deal announced between Nvidia and OpenAI in 2025 was reportedly never signed, and OpenAI scaled back chip purchases in June.

OpenAI gaming agent goals

▪Developer cozyblaze noted that GPT-6 Astra's Portal run aligns with OpenAI's 2016 technical goal to solve a wide variety of games using a single agent.
▪Developer cozyblaze characterized OpenAI's GPT-6 Astra as 'the worst model we'll ever get' while noting the run provides a glimpse of OpenAI's original vision.