Grok 4.7 was supposed to arrive around September 12 -- xAI's own count, by Elon Musk's September 2 post giving it "10 days." The window closed with no model card, no API identifier, and no announcement from xAI itself. Musk followed up on X the day before it expired, saying the model needs a few more days to cook -- and, unusually for this pattern of slips, said exactly why.
“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, September 11
The defect Musk describes is specific enough to be checkable once the model ships: a reinforcement learning setup that rewarded short answers so heavily the model started giving up on problems it could actually solve, rather than working through them at length. That's a different kind of failure than a capability or safety problem -- it's a tuning mistake in how the model was rewarded during training, and Musk frames it as fixable in days, not a redesign.
Every date Musk has given Grok 4.7, and what happened
- Jul 24, 2026 — Musk says Grok 4.7 ships "in 4 weeks"
- Aug 12, 2026 — Grok 4.6 ships instead -- itself weeks behind Musk's own July target -- and Musk narrows Grok 4.7 to "3 to 4 weeks"
- Sep 1, 2026 — Musk posts that Grok 4.7 "comes out in 10 days"
- Sep 11, 2026 — The 10-day window closes with no release; Musk cites a reinforcement-learning tuning problem and says the model needs "a few more days"
Grok 4.6 went through an almost identical pattern in July and August before it shipped -- announced, delayed twice, and confirmed only once it actually appeared in xAI's own release notes. The same standard applies to Grok 4.7: a shipped model needs a model card, an API identifier, and a price a developer can actually call, not a founder's post on X.
What's still unconfirmed is everything else Musk has claimed about the model: a 2.1-trillion-parameter count, a roughly 40% jump from Grok 4.6's 1.5 trillion, and training on additional SpaceX engineering and Starlink telemetry data. Grok 4.6 currently sits at 51 on the Artificial Analysis Intelligence Index, the independent measure this site tracks -- unmoved by anything Musk has said about 4.7, because none of it has been measured yet.
- Grok 4.7 missed its third target date on September 11, with no new date given.
- Musk says reinforcement learning penalized response length too heavily, making the model quit early.
- The 2.1-trillion-parameter figure and added SpaceX training data remain Musk's claims alone.
- Grok 4.6 itself shipped August 12, also weeks behind its original July target.
- Caveat: no model card, API identifier, or benchmark exists yet for Grok 4.7.