xAI / GrokElon's posts · 10 NovMUSK: AI STACK REWRITTEN IN C, PRE-TRAINING DONE

Grok 5 Should Exceed Opus 5.5, Maybe Within 2 or 3 Months

Musk says his team dropped JAX and SGLang, rewrote Grok's RL and inference stack in C/C++, finished pre-training and now has Grok 4.8 in RL with evals climbing.

Illustration of data streams flowing from a server hall up toward orbiting satellites
Musk says the new Grok training and inference stack is derived from the work behind Starlink. Illustration.Illustration
AT A GLANCE
Elon's posts · 10 Nov

These are Musk's own posts on X, which show what he said about Grok's progress but are not independent test results or benchmarks.

What changed
JAX/SGLang dropped; RL and inference stack rewritten from scratch in C/C++
Language mix
"C with a little C++, but not too much"
Training status
Training (aka pre-training) is done; SFT data needed a lot of cleanup
Current model
Grok 4.8 is in RL and evals are starting to climb
Next step
Soft rollout to Bot users later this week
STORY THREAD Grok 5 trainingFirst report in this thread. Updates will appear here.

Elon Musk pulled back the curtain on Grok’s engineering on 10 November with a burst of posts on X. The headline news: the team has dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++, a job he said took some extra time1. In the same post he said training, which he described as pre-training, is done1.

He followed with a look ahead to Grok 5, writing that it should exceed Opus 5.5 and that maybe that happens in 2 or 3 months2. And he tied the whole effort back to SpaceX’s satellite network, saying the new Grok stack is derived from work done by the engineers who created Starlink3.

What Musk said about the rebuild

The main post lays out the sequence plainly. The JAX/SGLang setup is gone, replaced by an RL and inference stack written from scratch in C/C++1. Pre-training is finished1. Then came the data work: Musk wrote that a lot of cleanup was needed on the SFT data to get that mostly right1.

He also gave a sense of the language mix. Replying to a user, he described the code as “C with a little C++, but not too much”4. In another reply, he made the case for that choice with a smile:

So easy to trigger Rust programmers 😂 Rust is a good language, but C is what you use for optimal performance on bare metal. With great power comes great responsibility.

That reply is his own explanation of the thinking: C, in his view, is what you use for optimal performance on bare metal5. In a separate reply in the same exchange, he added, “Real men manage their own memory”6.

Abstract illustration of a software foundation being rebuilt layer by layer
Musk says the RL and inference stack was rewritten from scratch in C/C++. Illustration.Illustration

Where Grok 4.8 stands right now

The model in the oven today is Grok 4.8. Musk wrote that 4.8 is in RL and the evals are starting to climb1. The next step is a soft rollout to Bot users later this week, with the team planning to see how it goes1.

He was refreshingly candid about the starting point. Grok 4.8, he wrote, “will bear no resemblance to 4.7 (thankfully)”1. He also said the model is not yet incredible, then explained why he is optimistic anyway:

It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special.

That is a builder’s update rather than a victory lap: a clear statement of where the model is, what was fixed underneath it, and what he expects from here1.

Grok 5 and the Opus 5.5 benchmark

The boldest line of the day came in a short reply. Musk wrote “Not as good,” then praised Opus 5.5 as “a masterpiece”2. He did not stop there:

His full words were that “Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months, with continued acceleration thereafter”2. The wording matters. He said Grok 5 “should” exceed it and that “maybe” that happens in two or three months, so ENN reads this as his expectation and a rough window, not a fixed launch date2.

It is also generous to a rival product. Calling another model a masterpiece while saying your own will pass it is a clear signal of how high Musk has set the bar for Grok 52.

The most striking thread in Musk’s posts is the link between Grok and SpaceX’s satellite network. He started with a point of pride about how SpaceX works: “At SpaceX, we write our own headers and calculate our own time (includes leap seconds)”3. He added that not many people will know what this means, but those that do will appreciate it3.

Then came the connection. Musk wrote that the new Grok AI/SI training and inference stack is derived from work done by amazing engineers to create Starlink3. In a follow-up post he went further on the fit between the two:

In that post he said the Starlink software, built for managing the movement of vast amounts of data through space and to the ground, “is also very well suited to training super intelligence”7. He framed it as a discovery, opening with “Turns out”7.

Starlink satellites
Musk says software built to move data through space and to the ground is very well suited to training super intelligence.Illustration

Put together, Musk’s posts describe two things: the new Grok stack is derived from Starlink engineering work3, and he sees Starlink’s software as well suited to training super intelligence7. He did not lay out further technical detail on how that work carried over, and ENN is not adding any.

How it fits the wider push

The Grok update landed on a busy day for Musk’s AI and energy ambitions. The same day, he posted that SpaceX has ordered 100 GW/year of HJT solar manufacturing equipment from Singulus in Germany, saying it is “for Starmind to create a sentient sun”8. ENN covered that order separately in SpaceX orders 100 GW a year of solar cell gear from Germany’s Singulus.

Read alongside that post, the Grok update shows a team willing to rebuild core software rather than stay on an existing setup. Musk himself acknowledged the cost in time, saying the rewrite “took some extra time”1. His bet is that getting the foundations right now pays off later1.

What the evidence shows, and what it does not

Everything in this article comes from Musk’s own posts on X on 10 November. They are a primary source for what he said: that the RL and inference stack was rewritten in C/C++1, that pre-training is done1, that Grok 4.8 is in RL with evals climbing1, and that he expects Grok 5 to exceed Opus 5.52.

The posts do not include eval scores, benchmark tables or a release date for Grok 5. His Grok 5 timing is phrased as “maybe” 2 or 3 months2, and the Grok 4.8 rollout is described as a soft rollout to Bot users to “see how it goes”1. ENN will report independent results when they appear.

What to watch next

  • The Grok 4.8 soft rollout. Musk said it goes to Bot users later this week1.
  • How 4.8 evolves. He said it is not yet incredible but expects it to evolve to be something special1.
  • Grok 5 against Opus 5.5. His window is maybe 2 or 3 months, with continued acceleration thereafter2.
  • More on the Starlink-derived stack. He has said the new stack comes from Starlink engineering work3 and that Starlink software is very well suited to training super intelligence7.

For now, the message from Musk is clear and confident: the RL and inference stack has been rebuilt in C/C++, pre-training is behind the team, and, in his words, this time the foundations are built right1.

Questions readers ask

What did Musk say changed in Grok's software?

He said the team dumped JAX/SGLang and rewrote the entire RL and inference stack from scratch in C/C++, which took some extra time. In a reply he described it as C with a little C++, but not too much.

Is Grok's pre-training finished?

Musk wrote that training, which he called pre-training, is done, and that a lot of cleanup was needed on the SFT data to get it mostly right.

When can people try Grok 4.8?

Musk said Grok 4.8 is in RL with evals starting to climb, and that there will be a soft rollout to Bot users later this week.

What did Musk say about Grok 5?

He wrote that Opus 5.5 is a masterpiece, but that Grok 5 should exceed it, maybe in 2 or 3 months, with continued acceleration thereafter.

How is Starlink involved?

Musk said the new Grok training and inference stack is derived from work done by the engineers who created Starlink, and that Starlink software is very well suited to training super intelligence.

Sources

Each story is checked against the original material and independent reporting. Numbers in the text point here.

  1. 1
  2. 2
    PrimaryMusk reply: Grok 5 should exceed Opus 5.5 X (@elonmusk) · Nov 10, 2026
  3. 3
    PrimaryMusk post: new Grok stack derived from Starlink work X (@elonmusk) · Nov 10, 2026
  4. 4
    PrimaryMusk reply: C with a little C++ X (@elonmusk) · Nov 10, 2026
  5. 5
    PrimaryMusk reply on Rust and C X (@elonmusk) · Nov 10, 2026
  6. 6
    PrimaryMusk reply on managing memory X (@elonmusk) · Nov 10, 2026
  7. 7
  8. 8