The short version

Grok 4.8 has not officially launched. Musk said in the early hours of November 18 (Taiwan time) that it has been soft-rolled-out to some Grok Bot users and needs another week or two of polish before general release. That is his own estimate. Specs, price and results do not exist yet.

When Grok 4.8 will come out

The timing rests only on Musk’s word. Counting from his posts in the early hours of November 18, “a week or two” lands around the end of November to early December.

Here is how the timing has unfolded. On November 10, Musk said a soft rollout to Bot users would come “later this week.” In the early hours of November 18 he first said 4.8 has not rolled out for general use and there are only trials behind the scenes, then about half an hour later said it has been soft-rolled-out to some Bot users but needs another week or two of polish before general release.

Now look at the pace of xAI (now SpaceXSI) itself. The official announcements came on July 16 for Grok 4.5, August 12 for 4.6 and September 21 for 4.7, gaps of 27 and 40 days. As of today it has been 58 days since 4.7, no announcement for 4.8 has appeared, and the newest model listed on the docs.x.ai models page is grok-4.7.

What Musk said

On November 10 (Taiwan time, early evening), a user asked on X for an update on 4.8, and Musk answered with one long post. His original words:

We have 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. 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.

A few technical terms, explained in plain language:

  • JAX: a math computation library made by Google, often used to train models.
  • SGLang: open-source software that lets a model answer quickly when a user sends a question.
  • Pre-training: the first stage, training a model from scratch to predict the next word.
  • SFT (supervised fine-tuning): teaching a model how to respond using data made of a question plus a reference answer.
  • RL (reinforcement learning): the system scores the model’s answers, and the model adjusts based on the scores.
  • Inference: the process of the model producing an answer after it receives a question.

Three more posts followed:

  • Later on November 10: someone asked how 4.8 compares with Claude Opus 5.5. Musk replied “Not as good. Opus 5.5 is a masterpiece,” and said Grok 5 should exceed Opus 5.5, possibly within two or three months (report by the tech outlet Decrypt).
  • Early on November 18, 01:56: “Grok 4.8 has not rolled out for general use yet, just some trials behind the scenes.”
  • Early on November 18, 02:29: “We did do a soft rollout to some Bot users, but still have another week or two of polish before general release.”

Earlier still, on October 30, he said 4.8 was taking longer than expected but would not be a disappointment. He said it is not being tuned for benchmarks, that he does not care how Grok scores on them, and that what matters is that it is incredibly useful.

How it differs from Grok 4.7

Every difference so far comes from Musk’s own words:

  • Stack rewrite: the whole RL and inference stack now runs on C/C++ instead of JAX and SGLang.
  • Musk’s verdict on 4.7: on September 14 he said 4.7 “should be” roughly on par with Opus 5.0 (not 5.1), better in some ways and worse in others, that multimodal performance (handling different kinds of content such as text and images at the same time) still needs fixing, and that 4.8 would be a “noticeable improvement.”
  • Where he places 4.8: he says it is not yet incredible and not as good as Opus 5.5, but that the foundations are built right this time.

4.7 was released on September 21 and reads up to 500,000 tokens at once. API pricing per million tokens is $2 input and $6 output for prompts under 200,000 tokens, and $4 input and $12 output at 200,000 or more. For earlier generations, see the complete Grok 4.6 guide and the Grok 4.5 guide.

Who gets it first

Musk said “some Bot users,” meaning users of Grok Bot. In the early hours of October 8 (Taiwan time) he said most questions people send to @Bot on X are simple and will be handled by a “lightning-fast version” of 4.8.

Grok Bot is included in the Premium, Super and Ultra plans of the X Subscription, and the Plus plan does not include it. For plans and pricing, see the Grok Bot guide. Musk’s posts do not mention a free tier or grok.com.

What is known and unknown

ConfirmedNot known yet
Pre-training is done (Musk, November 10)Official release date
RL is under way and eval scores are rising (Musk)Eval and benchmark numbers
Some Grok Bot users are in a soft rollout (Musk, early November 18)Which users, plans and regions
Another week or two before general release (Musk’s estimate)Whether it opens on grok.com, the X app or the API
The whole RL and inference stack was rewritten in C/C++Model ID, context length (how much it can read at once), price
Musk says 4.8 is not as good as Opus 5.5Feature list

Two other leaks are circulating online and are unconfirmed. In late September, one report said the string grok-4.8 appeared in xAI’s Grok Build codebase, and others say the name grok-4.8 once appeared in the model list of Cursor, the AI coding tool. Both rest on a single source and cannot be treated as specs or a launch time.

What to do now

  • Track progress: watch x.ai/news and the docs.x.ai models page. A 4.8 announcement with a model ID is what counts as an official launch.
  • Grok Bot users: wait for it to reach your account. You may be in the first batch, or you may have to wait for general release.
  • If your current tools work well: no need to stop for 4.8. For where Grok stands overall, see the complete Grok guide.
  • If you want to compare before deciding: wait for the official announcement, then choose based on price and independent evaluations.

This article will be rechecked and updated by November 25.