Nvidia CEO Claims AGI Arrival: Astra vs Fable 5.1 Game Test

On September 9th, 2026, Nvidia CEO Jensen Huang announced on X that AGI (Artificial General Intelligence) had arrived, congratulating the OpenAI team. He noted that OpenAI's new Astra model was trained on over 100,000 of his Grace Blackwell GPUs, with another 400,000 coming online soon. This claim sparked debate about whether Astra truly represents AGI and its implications for human obsolescence.

Astra vs. Fable 5.1: A Hands-On Comparison

To assess Astra's capabilities, a comparison was made between Astra and Fable 5.1, specifically focusing on game development, a field previously thought to be safe from AI due to the need for specific "feel and soul" and complex 3D graphics.

Game Development Capabilities

GPT-6 Astra has been dubbed a "Turbo AI machine god for 3D games." It can generate impressive games in 3JS, create detailed graphics in Blender, and integrate with game engines like Unity and Unreal. This suggests that game developer and 3D artist jobs may become increasingly rare.

Astra also achieved the top score on the ARC AGI benchmark, designed to test generalized intelligence rather than memorization. While this is impressive, it's worth noting that a team at Berkeley previously scored 99% on the same benchmark using Opus 4.8 and Fable 5, with the "harness" being the key differentiator.

Rocket Launch Simulator Challenge

To provide a practical comparison, both Astra and Fable 5.1 were given the same prompt: "Create a rocket launch simulator where you customize a rocket and launch it into orbit."

Astra's Performance

Astra completed the task in approximately 26 minutes. The resulting game featured highly detailed and beautiful 3D graphics with minimal errors, and a visually appealing UI. However, the UI design appeared to follow a common formula seen in other AI-generated games on platforms like Twitter. While a rocket could be launched into orbit, the gameplay itself was not described as particularly fun, with the most impressive aspect being the graphic of the rocket orbiting the globe.

Fable 5.1's Performance

Fable 5.1 took significantly longer to complete the task. The graphics were less detailed than Astra's, and the UI resembled older designs (e.g., Twitter Bootstrap). However, the gameplay was deemed superior. Fable offered more complex rocket customization options, incorporating scientific calculations that made it feel like a more legitimate simulator. The gameplay was also more engaging, with diverse possibilities for success or failure and cooler explosion animations.

User Preference

When presented to children, the 3-year-old preferred Astra's simpler game, while the 8-year-old favored Fable's more complex experience. This suggests that preference can be subjective, but Astra's ability to generate 3D graphics has dramatically improved.

Evolution of 3D Graphics Generation

A significant improvement in Astra's 3D graphic generation was observed. Just months prior, in July, a prompt for an exploded view of a mechanical watch resulted in "useless garbage." Rerunning the same prompt with Astra yielded a beautiful diagram that would have taken a human 3D designer hundreds of hours to produce.

This advancement raises concerns about a potential "new wave of slop" on platforms like YouTube, where high-effort 3D explainer videos (e.g., those by Jared Owen) could be easily replicated by AI, potentially diminishing the quality of online content.

The AGI Question: HorseTinder Test

To definitively answer whether Astra is AGI, it was tasked with building "HorseTinder" without any mistakes. After 21 minutes, Astra produced a design that was noticeably cleaner and more elegant than a similar design generated by Fable months ago. After extensive review, no mistakes were found in the generated code.

This flawless execution initially led to the conclusion that Jensen Huang was correct and AGI had been achieved. However, a minor imperfection—a few pixels being off on a carrot—revealed that it was a false alarm, and Astra is "just another dumb AI model."

Sponsor: Mobin

The article concludes with a promotion for Mobin, a tool for UI design. Mobin allows coding agents to study over 600,000 real UI screens and complete user flows from thousands of applications. Its MCP server integrates with tools like Claude Code, Cursor, and Codeex. Instead of generating generic designs, Mobin enables agents to gather references, identify design patterns, and adapt them to specific applications, with every result traceable to its original source to prevent design hallucinations. A 20% discount is offered for Mobin.