Pretraining finished, now in reinforcement learning
Elon Musk, who leads SpaceXSI (SpaceX Super Intelligence), the developer of the AI model "Grok," posted on X on Nov. 10 (the night of the same day Japan time) that the next model, "Grok 4.8," has finished pretraining and entered the reinforcement learning stage. He said it will begin rolling out gradually to users of the company's AI agent "Grok Bot" later this week to see how it does. The post was a reply to a user asking about the timing of Grok 4.8.
Pretraining is the stage in which a model's foundation is built from large amounts of data, and reinforcement learning is the stage in which the model is refined based on whether its answers are good or bad. Musk explained that it took extra time because the company dropped JAX, used for machine learning development, and SGLang, used for inference, and rewrote its reinforcement learning and inference systems from scratch in C/C++. The data for supervised fine-tuning (SFT), which teaches with example answers, also needed major cleanup, he said.
He said evaluation scores have begun to rise since reinforcement learning started, but did not give figures. He wrote that Grok 4.8 will be "nothing like" the previous version, Grok 4.7, and that while he cannot yet call it "great," he expects it to grow into something special because the foundation was built correctly this time.
"Not as good" as Opus 5.5
In a separate reply the same day, asked how Grok 4.8 compares with Anthropic's "Claude Opus 5.5," Musk answered that it is "not as good." Calling Opus 5.5 a "masterpiece," he said the next model, "Grok 5," should surpass Opus 5.5, and that this might be two to three months away.
The comparison is Musk's statement, and the company has not published benchmark results for Grok 4.8. He also did not give the timing of Grok 5 as a firm schedule.
Not yet listed on the official website
As of the morning of Nov. 12 (Japan time), Grok 4.8 was not listed on SpaceXSI's official website, its news list or its developer documentation (model list, release notes and pricing). The latest official model is Grok 4.7, announced on Sept. 21, and the page introducing the coding agent "Grok Build" also says it runs on the company's latest model, Grok 4.7. The release notes in the documentation were last updated on Nov. 2.
No official announcement has been found that the gradual rollout to Grok Bot users that Musk described as planned had begun as of the same time. Timing and pricing for API availability have not been given either.
On Oct. 30, Musk posted that Grok 4.8 was taking longer than expected and did not give a release date (previous article). In the Nov. 10 post, he gave "later this week" as the target for starting to offer it to Grok Bot users.
