Summary
Fireship’s Jeff runs a direct head-to-head between OpenAI’s newly released Astra (GPT-6) and Anthropic’s Fable 5.1, framed against Nvidia CEO Jensen Huang’s same-day declaration on X that Astra represents the arrival of AGI. To test the claim, Jeff gave both models the identical prompt — build a rocket launch simulator — at the same time. Astra completed the task in roughly 26 minutes, producing visually impressive 3D graphics with a clean UI; Fable finished later but delivered significantly deeper gameplay mechanics, more complex customization knobs, and physics-based calculations that made the simulation feel more substantive.
The video also highlights a notable 3D graphics leap in Astra: a Blender mechanical-watch diagram that GPT Soul failed to render correctly in July 2026 was rebuilt successfully by Astra just days later, a jump the creator calls alarming for the future of high-effort 3D explainer content on YouTube. A separate AGI litmus test — building a ‘HorseTinder’ clone without errors — took Astra 21 minutes and produced zero mistakes in a code review. The ARC-AGI benchmark context is discussed: Astra topped the leaderboard, though a Berkeley team had previously scored 99% using Opus 4.8 and Fable 5 with a custom harness. The video ends without a definitive AGI verdict but offers specific, reproducible output comparisons that illustrate where each model currently leads.
📺 Source: Fireship · Published September 09, 2026
🏷️ Format: Comparison
