Musk: Grok 5 Could Reach AGI

On Sept. 14 Elon Musk wrote on X that Grok 5 is the version he expects could reach artificial general intelligence, and that Grok 4.8, a 2.5-trillion-parameter model, would finish training within the week and enter reinforcement learning.

Musk outlined a short roadmap for Grok versions. He wrote that Grok 4.7 performs roughly on par with Anthropic’s Opus 5.0, that Grok 4.8 would be a noticeable improvement, that Grok 4.9 would likely reach Astra or Fable-class performance, and that “That will be Grok 5.” He added, “Grok 5 maybe better than anything. We shall see.”

According to Musk, Grok 4.8 runs on xAI’s new C++ software stack and has about 2.5 trillion parameters. xAI has used large compute clusters and data sources that include real-time information from X and engineering knowledge from SpaceX to scale its models.

The posts did not include technical benchmarks or firm release dates beyond the near-term training milestone for Grok 4.8.

Experts have raised concerns about models that approach human-level performance, naming risks such as misaligned goals, difficulty maintaining control as systems become more capable, sudden capability jumps, job displacement, concentration of power, and misuse by malicious actors.

On the same day Musk backed a public call from Anthropic executive Dario Amodei urging the industry to slow the pace of capability development.

xAI’s statements set expectations for further model upgrades in the near term while leaving outcomes and timing uncertain.