← Back to live feed1 story
The GPT-6 Astra model overcame the automation game Factorio with enemies active after nearly 44 hours of simulation time. The run required 4 days 11 hours of total processing on a Codex goal clock and cost roughly $4,500 in API fees. Costs could decrease 50 times if the system offloads tasks to Luna or utilizes standard scripts for common operations.
The system failed to solve embodied tasks, though reports suggest the result indicates robotics may eventually adopt similar scaling laws. Future benchmarks include using Astra to coach other AI models, such as Muse Spark or Qwen 3.8 27B, to solve random seed challenges.
Sign in to suggest edits
Key sources
- SOURCE@mattparlmer“At API pricing, logs indicate it would've cost about $4500”x.com
- SUPPORT@steipete“GPT-6 Astra has just beaten a standard game of Factorio with enemies enabled after nearly 44 hours in-game time”x.com
- SUPPORT@chris_j_paxton“GPT-6 Astra did not solve embodied tasks. But it suggests a more consequential possibility: robotics may inherit the scaling curve”x.com