SpaceXAI starts a soft rollout of Grok 4.8, with no model card or benchmarks


SpaceXAI, the AI company formerly known as xAI, has started releasing Grok 4.8, its next large language model, to a limited group of users. It has not published a model card, benchmark results, API details or pricing, which leaves teams that vet models before using them with little to work from.
A note on the name: SpaceX acquired xAI in February 2026, and the company officially renamed itself SpaceXAI in July. Many reports still call it xAI, but SpaceXAI is the name it now uses, including on its developer documentation. Grok is still the name of its AI models and chatbot.
What SpaceXAI has said
On 10 November, Elon Musk said that pre-training was complete and the model was in reinforcement learning, with evaluation scores “starting to climb”. He added that the company had dropped its previous tools, JAX and SGLang, and rewritten its entire reinforcement-learning and inference stack from scratch in C/C++. He said a soft rollout to “Bot users” would follow later that week, and cautioned that the model “is not yet incredible”.
By 13 November the official Grok account confirmed that the soft rollout to Bot users was under way.
What is still missing
SpaceXAI has published no documentation of the model’s capabilities, limits or safety testing. Musk said evaluation scores were “starting to climb” but gave no figures.
Developers can’t try it yet either. There is no API model ID, and the company’s models page still lists Grok 4.7 as its newest model, with no mention of 4.8. Neither post explains who counts as a “Bot user” or whether everyone in that group gets the model at once, and there is no date for general availability.
Why the model card matters
A model card is where a provider sets out what a model was tested for, where it falls short and how it handles harmful requests. Grok 4.8 is reaching users before that paperwork exists. Security and risk teams have nothing to judge how it treats sensitive data, how easily it can be manipulated, or how it compares with the models they already use.
The rollout could also outpace any review. If it widens to the Grok accounts your staff already use, they may start using the new model before anyone has assessed it, and you can’t test it in a controlled way until an API model ID exists.
What security teams can do
Nothing is urgent yet. When Grok 4.8 reaches wider use:
- Hold off approving it for work until SpaceXAI publishes a model card and benchmark results.
- Once an API model ID exists, test it against your own use cases, including prompt-injection and data-handling checks.
- Remind staff not to paste confidential or personal data into unapproved AI tools, and that this includes new model versions in apps they already use.
- Keep a list of approved models and versions, so a new release doesn’t slip into use by default.