Where Grok 4.8 stands
● Pre-training is finished — Musk confirmed on 10 November that training is done and the model is now in reinforcement learning.
● The whole stack was rewritten — xAI dropped JAX and SGLang and rebuilt the RL and inference stack from scratch in C and C++.
● Limited rollout to Bot users this week — No date given for a wider public release.
● Musk rates Opus 5.5 above it — He called Anthropic model a masterpiece and said Grok 5 should pass it in two or three months.
Elon Musk gave a status update on Grok 4.8 on 10 November 2026, and the interesting part is not the model — it is what xAI tore out to build it.
They threw away the toolchain
In Musk’s words: “We have dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, which took some extra time.” JAX and SGLang are standard infrastructure for training and serving large models. Replacing both with hand-written C is not a tweak.
He described the result as “C with a little C++, but not too much” with some assembly, and added that the team writes its own headers and calculates its own time including leap seconds. His summary: “This is tight code.”
The stated cost of that decision is schedule. Musk’s mid-September roadmap had Grok 4.8 finishing training that week and moving into reinforcement learning, and it is reaching that point in November.
Current status, in his words
- “Training (aka pre-training) is done.” He added that a lot of cleanup was needed on the SFT data to get it mostly right.
- “Now, 4.8 is in RL and the evals are starting to climb.”
- A limited Bot-user rollout later that week, to see how it goes. No wider date.
- The model is 2.5 trillion parameters trained with the new C++ stack, per the September roadmap.
- “Will bear no resemblance to 4.7 (thankfully)” — not yet incredible, but the foundations are built right.
The comparison Musk made himself
Asked how Grok 4.8 compares with Claude Opus 5.5, Musk said it was not as good and called Opus 5.5 a “masterpiece”. He said Grok 5 should exceed it, and that this might happen in two or three months.
A chief executive naming a competitor’s model as better than his own unreleased one is rare enough to be worth noting. It also sets a public marker: Grok 5, past Opus 5.5, by roughly February.
What it means if you use Grok today
- Nothing changes this week unless you are a Bot user in the limited rollout.
- The rewrite is the thing to watch, not the version number. A hand-built C inference stack is how you get cheaper serving, and serving cost is exactly what Musk cited two weeks ago when Grok Bot usage limits tightened. We covered that and the plan limits here.
- Treat the timeline as a direction, not a date. The September roadmap put this point in September.
Frequently asked questions
When is Grok 4.8 coming out?
A limited rollout to Bot users was expected in the week of 10 November 2026. Musk gave no date for a wider public release.
What changed technically in Grok 4.8?
xAI dropped JAX and SGLang and rewrote the entire reinforcement learning and inference stack from scratch in C and C++, with some assembly. Musk said the model is 2.5 trillion parameters and that pre-training is complete.
Is Grok 4.8 better than Claude Opus 5.5?
Musk says no. Asked directly, he said Grok 4.8 was not as good and called Opus 5.5 a masterpiece, adding that Grok 5 should exceed it in perhaps two or three months.
Why rewrite the stack in C?
Musk did not give a reason beyond calling the result tight code, but a hand-built inference stack is the usual route to lower serving cost. xAI cited serving costs when Grok Bot usage limits tightened at the start of November.