Twitter user cozyblaze got GPT-6 Astra to play through the original Portal from start to finish, watching credits roll after 23 hours 42 minutes. The experiment started on September 4 at 5 PM and ended September 5 at 4:43 PM. Astra ran in maximum reasoning mode via MCP and a modified SourcePauseTool: the game would pause, the model would receive a screenshot, character coordinates, and camera angle, plan its next move, send commands, and unpause. This cycle repeated for nearly 24 hours.

Pure gameplay without reasoning pauses amounted to roughly 2 hours (YouTube recording). According to the published log, the run consumed 434.8 million tokens total: 433.2 million input and 1.6 million output, with 1.18 million used for reasoning. The model made 3,300 tool calls, 3,210 of which were in-game commands.

The key economics: 426 million of the 433 million input tokens hit cache, bringing the API cost to approximately $574 at published rates ($10/M input, $50/M output, $1/M cached input). Without caching, the same run would have cost around $4,400. In practice, cozyblaze spent nothing beyond a $200/month Codex subscription.

cozyblaze on X

Related: OpenAI Launches GPT-6 Astra with Strong Benchmark Results, OpenAI Publishes Detailed API Guide for GPT-6 Astra, OpenAI’s Astra Solves 10 Longstanding Math Problems, Cortical Labs Human Neurons Play Doom Faster Than GPT-4