Musk says Grok 4.8 has finished pre-training and will reach some Bot users later this week
The post Alex Diaz quotes says SpaceXSI dropped JAX and SGLang and rewrote reinforcement learning and inference in C and C++. There is no score and no API name. A separate reply, in his account, puts the model behind Opus 5.5 and Grok 5 two or three months out.
Feature3 minIn-depth
James Okonkwo3 min read
- Grok
- SpaceXSI
- Grok Bot
- Training

Photo: ArtificialWatchSource page
Elon Musk said on Tuesday that Grok 4.8 has finished pre-training, is in reinforcement learning, and will get a limited release to Bot users later this week. Alex Diaz at TeslaNorth filed on the post the same day and embedded it. ArtificialWatch times the post at 09:35 UTC and says it checked the wording against the original. Diaz refers to the lab as SpaceXSI. Neither account prints a benchmark, a price, or a model id.
The sentences both accounts carry are specific about the software and vague about the result. Musk wrote that the team has “dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.” “Training (aka pre-training) is done.” “Lot of cleanup was needed on the SFT data to get that mostly right.” “Now, 4.8 is in RL and the evals are starting to climb.” “We will do a soft rollout to Bot users later this week and see how it goes.” He does not name a day inside the week, a share of Bot users, or a number behind “starting to climb.” Supervised fine-tuning is the step the “SFT” refers to. Reinforcement learning is the step he says is underway now.
Grok 4.8 will bear no resemblance to 4.7 (thankfully). It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.
— Elon Musk on X, Nov 10, 2026, as quoted by TeslaNorth
Replies about the code, and one about Opus
Diaz reports further replies from the same day, after the training note. Musk described the new code as “C with a little C++, but not too much,” and said some of it is assembly. Diaz writes that Musk tied the software to engineers who built Starlink, and quotes him: “we write our own headers and calculate our own time (includes leap seconds).” Another line Diaz prints: “One day, super intelligence will write a better software stack, but that day has not yet come.” Those replies are about who wrote the stack. They are not a description of what 4.8 scores.
In a separate reply, Diaz writes, Musk was asked how 4.8 would compare with Anthropic’s Opus 5.5. On Diaz’s account he said it was “not as good,” called Opus 5.5 a “masterpiece,” and said Grok 5 “should exceed Opus 5.5 and maybe that happens in 2 or 3 months, with continued acceleration thereafter.” The two or three months are attached to Grok 5, not to a public 4.8. Diaz writes that Musk did not say when a wider release of 4.8 might come. The rollout he did describe is to Bot users, the agent, not a line on an API price card.
What the September note contained, and what Tuesday’s does not
Diaz sets the post against the roadmap Musk gave in mid-September, when he said 4.8 would finish training that week and start reinforcement learning, and described it as a 2.5-trillion-parameter model on a new C++ stack. Tuesday’s post does not restate a parameter count. The pre-training stage he had placed in that September week is the stage he now says is finished. The reinforcement-learning stage is the one he says has begun. ArtificialWatch adds a check the post does not: at 10:51 UTC, it writes, OpenRouter’s newest xAI model was still grok-4.7. The soft rollout he described is the one he said the lab would “see how it goes” on. Nothing published Tuesday gives a developer a model id to call.


