- @elonmusk (circa 2026-09-11) stated that Grok 4.7 "still needs a few more days to cook."
- Attributed to RL potentially over-penalizing response length: giving up too early on difficult problems and self-review not being rigorous enough.
- Grok 4.7 was previously reported to have around 2.1 trillion parameters, compared to Grok 4.6's approximately 1.5 trillion; the current flagship is still version 4.6.
- There were earlier rumors of a release window around September 12th; this update pushes back the public availability date further.
Elon Musk replied on X, stating that xAI's next-generation flagship model Grok 4.7 "still needs a few more days to cook" and is not yet ready for official release.
He explained that the team might have over-penalized response length during the reinforcement learning (RL) phase, causing the model to give up on difficult tasks prematurely even though it is capable of completing them. Additionally, the model's self-review of its answers has not yet reached a sufficient level of rigor.
Grok 4.7 needs a few more days to cook.
We might have penalized response length too much (or something) in RL, as it still gives up on hard tasks (that it can do!) too early and isn’t yet sufficiently rigorous in checking its work.— Elon Musk (@elonmusk)
Grok 4.7 was previously highlighted by Musk as having around 2.1 trillion parameters in its new pre-training scale (compared to Grok 4.6's approximately 1.5 trillion), and it was claimed to outperform its predecessor in most aspects, though service latency might be slightly higher. The training also incorporated SpaceX engineering-related data to enhance real-world engineering reasoning. The previously rumored release window had been around September 12th, but this "few more days" update pushes back the public availability further.
This update is a product timing announcement rather than a new feature list: the core issue is framed around RL trade-offs — how long the model is willing to persist on difficult problems and how rigorously it reviews its own output — rather than hardware bottlenecks or missing components. xAI has not yet released an official model card or benchmark report; the current flagship is still Grok 4.6, which was released on August 12th.
Primary source: @elonmusk 2026-09-11 reply (original text above). Parameter and training background references are based on Musk's earlier public statements and secondary reports for verification.