Musk Says Grok 4.8 Will Start a Limited Rollout to Grok Bot Users This Week

In a post on X, Musk said pre-training was complete and the model was now in reinforcement learning, with evaluation scores starting to improve. He gave no day for the release. He also did not say how many users would get access, whether developers on xAI's API would be included, or what it would cost. xAI had not published an announcement, model card or benchmark results for Grok 4.8 as of Tuesday, according to a release tracker kept by CellCog, an AI company that follows the model.

Musk blamed part of the delay on an engineering overhaul. xAI had dropped JAX and SGLang, open-source tools widely used to train and serve large models, and rewritten its reinforcement learning and inference software from scratch in C and C++. He said the team also had to clean up the data used for supervised fine-tuning. He described the model as not yet great but substantially different from Grok 4.7, according to accounts of the post.

The release will show whether xAI can keep pace with Anthropic, OpenAI and Google, which have all shipped new frontier models in recent weeks. Grok 4.7 came out on September 21 at $2 per million input tokens and $6 per million output tokens, the same price as Grok 4.6, according to CellCog. In September Musk described Grok 4.8 as a 2.5 trillion-parameter model trained on the new software stack. xAI has not confirmed parameter counts for any Grok 4 model.

Musk has been resetting expectations for weeks. On October 30 he acknowledged that the model had taken longer than expected, CellCog's tracker shows. When a user asked on Tuesday how Grok 4.8 compared with Anthropic's Claude Opus 5.5, Musk replied that it was not as good. He predicted that Grok 5 would overtake Opus 5.5 within two or three months, according to CellCog.

The soft rollout will give users their first look at the model. Musk has said most Grok Bot requests will eventually be served by a fast version of Grok 4.8.