JaySym_AI: "Elon Rebuilt Grok From Scratch in C and 4.8 Lands This Week". Facts on this page cite the primary sources below, not the video.

The claim: Grok 4.8 is out this week, rebuilt from scratch.

The proof: What's on the record is two replies from Elon Musk on Tuesday, Nov. 10. At 4:35 AM ET he wrote that SpaceX SI has "dumped JAX/SGLang and rewritten the entire RL and inference stack from scratch in C/C++," that pre-training "is done," that "4.8 is in RL and the evals are starting to climb," and: "We will do a soft rollout to Bot users later this week and see how it goes." Less than an hour later, asked how Grok 4.8 compares with Anthropic's Claude Opus 5.5, he answered: "Not as good. Opus 5.5 is a masterpiece." There's no release note, model card or benchmark from SpaceX SI yet. Its news page still lists Grok 4.7 as the latest model.

Editorial Thesis

This is a roadmap post from the CEO, not a launch. It's also unusually candid. Musk didn't call 4.8 the best model in the world. He said it's "not yet incredible," that it's behind Opus 5.5, and that the real bet is the new software underneath it. The useful question for Grok Bot users isn't "is 4.8 better?" It's "how will I know when I'm on it, and what changes when I am?"

Honest Verdict

  • Confirmed (Musk's posts, Nov. 10): pre-training done; 4.8 in reinforcement learning (RL); RL and inference stack rewritten in C/C++ after dropping JAX and SGLang; a soft rollout to Grok Bot users planned "later this week"; Musk rates it below Claude Opus 5.5.
  • Not published: a date, which plans or countries get it first, benchmark numbers, pricing, API availability, or a model card.
  • Prediction, not fact: "Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months."

What Musk actually said

When (ET)PostWhat it says
Mon Nov. 9, 7:23 PMUser question"Can we have an update on when to expect Grok 4.8 please"
Tue Nov. 10, 4:35 AMMusk replyJAX/SGLang dropped; RL + inference stack rewritten in C/C++; pre-training done; SFT data cleaned up; 4.8 in RL, "evals are starting to climb"; soft rollout to Bot users "later this week"; "will bear no resemblance to 4.7 (thankfully)"; "not yet incredible"
Tue Nov. 10, 5:22 AMMusk replyOn Opus 5.5: "Not as good. Opus 5.5 is a masterpiece. But Grok 5 should exceed Opus 5.5 and maybe that happens in 2 or 3 months"

The first reply had about 3.5 million views by Wednesday morning.

What the jargon means

  • JAX and SGLang are widely used open-source tools for training and serving AI models. Musk says SpaceX SI replaced them with its own code written in C and C++.
  • Pre-training is the big first pass where a model learns from huge amounts of text. SFT (supervised fine-tuning) teaches it from curated examples. RL (reinforcement learning) is the later stage that rewards better answers. Musk says 4.8 is in that last stage now.
  • Inference is running the model when you use it. A faster inference stack can mean quicker or cheaper answers. Nobody has published numbers showing that yet.

Why "soft rollout to Bot users" matters

Musk named Grok Bot, SpaceX SI's agent product, as the first place 4.8 shows up. He didn't mention the Grok app, the API or Tesla cars. A soft rollout usually means some users first, watched closely, then wider. So some Bot users may see different behavior this week while others don't, and SpaceX SI may not announce exactly when it switches.

What to check if you use Grok Bot

  1. Watch the official Grok Bot changelog and @bot for a line that names the model.
  2. Re-run one task you know well and compare the result and the usage it takes. Don't rely on someone else's demo.
  3. Expect rough edges. Musk himself called it "not yet incredible."
  4. If you build on the Grok API, nothing in these posts says 4.8 is coming there this week.

What this does not prove

  • It does not prove Grok 4.8 will ship this week. "Later this week" is a plan.
  • It does not prove 4.8 is better than Grok 4.7 at any task. No evals were published.
  • It does not prove Grok 5 will beat Opus 5.5. Musk said "maybe" and "should."
  • It does not mean the model was "rebuilt from scratch." Musk said the RL and inference stack was rewritten.

Bottom Line

On Nov. 10, Musk said Grok 4.8 is done pre-training, is in RL on a C/C++ stack SpaceX SI wrote itself, and gets a soft rollout to Grok Bot users later this week. He also said it's not as good as Claude Opus 5.5. Until SpaceX SI publishes a release note or numbers, treat it as a CEO roadmap, and test it yourself when it reaches you.

Source notes

Both Musk replies read on X on Nov. 11 at about 5 AM ET (view counts approximate; times converted from UTC to ET). SpaceX SI news page read the same morning (latest model launch: Grok 4.7). Tech City Authority (Nov. 10) for the absence of an official announcement. Video: JaySym_AI's Nov. 10 explainer; its "rebuilt from scratch" title is the creator's framing.

Sources