Musk: Grok 5 Has 10% Chance of Achieving AGI at Barron's

At the Barron’s Investment Conference, Elon Musk offered his first public estimate on artificial general intelligence (AGI), stating that Grok 5—slated for release in Q1 2026—has “about a 10% chance” of achieving AGI.

Grok 5 will feature a 6-trillion-parameter sparse Mixture-of-Experts (MoE) architecture, using roughly 1.8× the memory of GPT-4o. It will process text, images, audio, and real-time video with frame-level latency under 120 ms. Training data comes from X (formerly Twitter), including 500 million daily public posts and 200 million hours of video streams, fed into the model in real time per user agreements.

The model is being trained on xAI’s Memphis supercomputer—a cluster of 100,000 NVIDIA H100 GPUs delivering 750 petaflops of peak performance—with pre-training expected to wrap up by February 2026.

Musk defines AGI as the ability to understand and reason over real-time video like a human—for example, performing live car repairs or interpreting complex street scenes. In internal tests, Grok 5 passed 68% of 300 visual reasoning tasks, nearing the human baseline of 75%. During a live demo, it analyzed 30 seconds of drone footage in real time—identifying 47 object types, estimating traffic flow, plotting the fastest route, and flagging construction hazards—without any task-specific fine-tuning.

Musk warned that AGI could arrive by 2030 and called for a globally verifiable, “disarm all humanity”–style cooperative framework to ensure safety. xAI has dedicated 10% of its computing capacity to alignment and safety research.