Musk Says Grok 4.8 Test Is Planned for This Week After Software Rewrite

What Musk announced

November 11, 2026, 10:38 a.m. Central.

Elon Musk says Grok 4.8 has finished its initial training and could reach Grok Bot users in a limited test later this week, after a software rewrite he says slowed development.

The remarks came Tuesday in a reply on X to an account named @cradrrat, which asked when to expect the model. Musk said pretraining was complete and the model was undergoing reinforcement learning, a later phase that refines behavior using feedback or rewards. He said evaluation results were improving, but published no scores.

His proposed “soft rollout” is a plan, not confirmation the model is already available. He did not give an exact day, the number of participants, or a date for a wider release.

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.
— Elon Musk (@elonmusk) November 10, 2026

X did not permit direct access during Newstro’s check. The reply and the question it answered were checked against an independent archive and contemporaneous reporting by Supercharged Daily and Tesla North.

What the rewrite means

Musk said the team replaced JAX and SGLang with software written in C and C++ for reinforcement learning and inference — the process of generating a model’s response to a prompt. He also said the team had to clean up the data used for supervised fine-tuning.

JAX is a tool for high-performance numerical and machine-learning work; SGLang is a framework for serving language models. A change in the underlying tools can affect how a system is built and operated, but Musk provided no measurements demonstrating a particular speed, cost or quality gain from the rewrite.

In another exchange reported by Tesla North, he credited work by Starlink engineers with helping shape the new stack. Those are his descriptions of the development work, not independently verified performance results.

A candid comparison — without scores

In a separate Tuesday reply, @Dimdv99 asked how Grok 4.8 would compare with Anthropic’s Claude Opus 5.5. Musk answered that it would be “not as good” and praised the rival model, according to the archived exchange and Decrypt’s reporting.

He also predicted that a later Grok 5 could surpass Opus 5.5 in two or three months. That is a projection, not a release date or a measured result. Musk did not provide head-to-head Grok 4.8 scores with either statement.

What users can check

As of this reporting, SpaceXSI’s public model documentation listed Grok 4.7 as its latest model and did not list Grok 4.8. That does not rule out private testing; it means the promised Bot rollout should not be confused with general API availability.

The company introduced Grok 4.7 on September 21. In mid-September, Musk had said he expected 4.8’s initial training to finish that week, according to Supercharged Daily. Tuesday’s update came weeks after that earlier target.

The new model’s broader availability, price, independent test results and specific Bot-user eligibility remain unanswered. Newstro has previously reported on Grok Bot usage complaints and efficiency work; the 4.8 plan concerns a model update, not proof those service issues are resolved.

The timeline

Dates below are Central time unless noted. September 13, 2026: Musk said 4.8’s initial training would finish that week, according to Supercharged Daily. September 21: SpaceXSI introduced Grok 4.7. November 10, 3:35 a.m.: Musk said 4.8 pretraining was done and outlined a Bot test later that week. November 10, 4:22 a.m.: he offered his comparison with Opus 5.5. The test remains a stated plan.