Elon Musk has indicated on X that the next-generation model Grok 4.8 is a roughly 2.5 trillion parameter (2.5T) model trained using the company's new C++ software stack. He expects it to complete training this week and begin reinforcement learning (RL).
Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL
— Elon Musk（@elonmusk）
In the same conversation, when asked how close 4.8 is to AGI, Musk simply replied: "That will be Grok 5," pushing the AGI-level target to a later version number.
He later provided a longer response to align the entire product hierarchy with the public narrative:
Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance.
Grok 4.8 will be a noticeable improvement.
Grok 4.9 is probably Astra/Fable class.
Grok 5 maybe better than anything. We shall see.— Elon Musk（@elonmusk）
In other words, in his narrative, Grok 4.7 is positioned as roughly on par with Anthropic's Opus 5.0 (not 5.1), with the need to improve multimodal performance; 4.8 is the next tier of "noticeable improvement"; 4.9 is placed at the Astra/Fable level; while Grok 5 remains open to the possibility of "maybe better than anything." The current flagship model publicly released is still Grok 4.6 from August 12th; Musk previously indicated that 4.7 still needed a few more days to "cook," and the company has not yet released the full model card or benchmark report.
This update pertains to the training schedule and competitive positioning rather than a product launch specification sheet. The key new information is that 4.8 is the first version to be explicitly linked to the new C++ training stack, and the timeframe for "completing training → starting RL" is set for this week. The actual launch date, delays, and pricing will still depend on subsequent product announcements from xAI/SpaceXAI.