Elon Musk has already written the spec sheet. He just wrote it on X.
Until xAI puts a card on a domain, those posts are the primary source. Pretending they are not information is a different kind of dodge. The honest move is to file them as claims, with dates, and to keep the shipped model in a separate pile so the two do not bleed.
On 25 July 2026 he posted the cadence in one line: Grok 4.6 in two weeks, Grok 4.7 in four. On 28 July he named the sizes. “Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL.” Then: “Grok 4.7 will be the 2.1T model released a few weeks later. This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better token efficiency.”
That last sentence is the actual product claim. He is not promising a faster box. He is promising more work per token, paid for in latency. If it holds, a larger model can still be cheaper per finished task when tokens-per-second drop. If it does not, you bought a slower rumor. There is no independent tokens-per-second number, and no per-task cost, attached to 4.7. There is only that line.
4.6 actually arrived. 12 August, five days after “around 7 August,” close enough that the cadence still looks like a plan. The note on x.ai is a real page. The docs list grok-4.6. Context is 500,000 tokens. Price starts at $2 per million input tokens and $6 per million output. On the Artificial Analysis Intelligence Index it scores 61, tying GPT-5.6 Sol, one point under Fable 5 Max at 62, up from Grok 4.5 High at 56. Knowledge cuts off on 1 February 2026. That pile is not rumor. You can curl it. You can hate the evals in public.
The 4.7 pile, as of 22 August, is still Musk. After 4.6 shipped he moved the window to three to four weeks, which is early or mid September if the clock holds. Reporting around the same day says initial training is done and the model is in supplemental training on SpaceX engineering data. He wrote that the corpus is “so awesome and unique” he would be shocked if any model were better at real-world engineering than 4.7. He also said Anthropic is a great company and will probably release improved models soon. Competitive talk is not an eval.
What he has not said, and what xAI has not posted: an API identifier, a price, a context window, a SWE number, a tokens-per-second figure for “slightly slower,” dense versus mixture-of-experts, whether vision comes along. Grok 4.5 already sits at 500k context and the same $2/$6. People will assume 4.7 inherits both. Assumption is not a spec.
Grok 5 is a different file. Do not stack it onto 4.7 to make 4.7 feel heavier. 4.7 is a point release with a founder quote. 5 is the named jump that has been “in training” while the 4.x line kept shipping.
On 6 January 2026, xAI said it in a clause on its own site, in the $20 billion Series E note: “Looking ahead, Grok 5 is currently in training.” That is the company, not a reply guy. Reuters ran it the same day. The cluster attached to the story is Colossus 2 in Memphis. On 17 January Musk posted that the Colossus 2 supercomputer was operational, “first Gigawatt training cluster in the world,” with talk of a 1.5GW upgrade in April.
The size talk is older and softer. At Baron Capital in November 2025 he discussed a target around 6 trillion parameters, multimodal across text, images, video, and audio. Treat 6T as a founder target, not a measured card. A 10-trillion variant shows up in community reporting. That one is not in the Series E note. Leave it in the drawer.
The calendar for 5 is a graveyard. On 7 August 2025 he said it would arrive “before the end of this year and it will be crushingly good.” It did not. Q1 2026 went by. Q2 went by. The 4.x point releases filled the silence. During SpaceX’s August 2026 earnings call he put 5 before the end of 2026 again, and said it would take about 25 years of SpaceX data. File the distinction: 4.7’s SpaceX story is supplemental training. 5’s is the archive.
He has also floated a 10 percent and rising chance that Grok 5 is AGI. No test. No threshold. Put it next to crushingly good.
So the roundup, labeled. Musk says 4.7 is 2.1 trillion, slower to serve, better token efficiency, better at engineering because of SpaceX data, here in September if three-to-four weeks from 12 August is real. xAI has not repeated any of that on a product page. Musk and xAI say 5 is already training on a gigawatt cluster, aimed at 6 trillion, multimodal, year-end if this clock holds after the last two did not, fed on a quarter-century of rocket paperwork. 4.6 is the one with a name the API will take. 4.7 is the one you can quote. 5 is the one that has been coming since last August.
I will update this when an identifier appears. Until then the lantern stays on 4.6, and the rumors stay labeled.