Musk publishes xAI Grok roadmap: Grok 4.8 finishes 2.5T-param training this week, Grok 5 targets AGI

On September 14, 2026, Elon Musk posted xAI's full Grok model roadmap on X and identified Grok 5 as the candidate xAI expects to reach AGI. Musk confirmed Grok 4.8 uses 2.5 trillion parameters, a roughly 67 percent increase over Grok 4.6, trained on a brand-new C++ training stack developed in-house, with pretraining wrapping this week before the model moves into reinforcement learning. Earlier Grok releases were trained on a Jax-based stack; 4.8 is the first to switch to the C++ implementation, and a planned 3-trillion-parameter run will continue on the new stack with improved data quality. On the capability ladder, Musk said Grok 4.7 is roughly on par with Anthropic Opus 5.0, stronger in some respects and weaker in others, with multimodal performance still needing work; Grok 4.8 will be a noticeable improvement; Grok 4.9 targets the GPT-6 Astra / Fable class; and Grok 5 may be better than anything else available. Asked how close Grok 4.8 gets to AGI, Musk answered that the AGI goal is reserved for Grok 5, noting xAI has closed the gap with established labs quickly but that matching or exceeding human-level performance across all tasks remains a first-of-its-kind bar for any AI company. On the same day, Musk publicly endorsed Amodei and Altman's calls to pace frontier AI development, writing that peer review of AI by competitors is the right way to start coordinated governance, a stance that sits in tension with his own public commitment to an AGI-grade Grok 5. Chip stocks moved sharply on the day: Nvidia fell 3.46 percent and AMD dropped 6.24 percent as traders priced in the possibility that AI safety concerns and pacing calls could weigh on near-term AI hardware demand.