I built the same game with Astra and Fable 5.1... only one was fun

Summary

The video examines whether OpenAI's GPT-6 Astra represents true artificial general intelligence following an announcement by Nvidia CEO Jensen Huang, who noted that the model was trained on over one hundred thousand Grace Blackwell NVLink72 GPUs with four hundred thousand more coming online. While Astra achieved leading scores on the ARC-AGI-3 benchmark, human evaluations and real-world coding tests provide a more grounded perspective on what frontier models can actually deliver.

To evaluate Astra against Anthropic's Fable 5.1, both models received an identical prompt requesting a 3D rocket flight simulator built with HTML, JavaScript, Three.js, and relevant physics libraries. Astra finished its implementation in twenty-five minutes and fifty-two seconds. The resulting game, Orbital Lab, featured high-fidelity 3D assets, planetary orbital views, and a polished user interface. However, the UI shared a noticeable formula with other AI-generated web games, and the actual mechanics lacked depth and challenge.

In contrast, Fable 5.1 took significantly longer to complete its generation and warned that it had written custom physics rather than pulling in external rigid-body libraries. The visual presentation of Fable's project resembled early Twitter Bootstrap styling with rudimentary 3D textures. Despite the underwhelming aesthetics, Fable delivered superior gameplay featuring detailed rocket component customization, scientific flight calculations, atmospheric dynamics, and entertaining explosion animations. Testing the two games with children yielded a split outcome, as a three-year-old favored the visual simplicity of Astra while an eight-year-old preferred the explosive mechanics of Fable.

Astra also showcased major improvements in pure 3D rendering. When tasked with producing an exploded view of a mechanical Rolex Submariner watch, Astra generated an intricate, accurate technical diagram in minutes. This contrasted sharply with an earlier attempt from GPT Sol months prior, which had produced broken geometry. While such automated rendering saves hundreds of hours compared to manual creation in tools like Blender, it also lowers the barrier to mass-produced 3D video content.

Finally, Astra was tested in ChatGPT Work with a prompt to build a functional web application named Horse Tinder without making mistakes. The model completed the build in twenty-one minutes, producing an elegant, fully interactive interface with swipeable cards, detailed profiles, and functional filters. A rigorous line-by-line inspection of the source code revealed no bugs or syntax errors, establishing that frontier models have made substantial strides in autonomous software construction despite remaining short of true generalized autonomy.