On Thursday, September 24, 2026, Elon Musk put a clock on one of the boldest promises in tech. In two short posts on X, he said SpaceXAI, the AI arm of SpaceX, could match the best models from Anthropic and OpenAI within a few months and take the lead around the six month mark.
Markets barely reacted. Shares of SpaceX rose just 0.2% overnight after Musk shared his cautiously optimistic view that a Claude Fable and GPT-6 class model was two to three months away. Investors seem to want proof, not promises.
So I did what a promise like this deserves. I pulled 12 specific claims out of Musk's posts and the coverage around them, then checked each one against benchmark data, SpaceX's own financial filings and reporting from established outlets. Some hold up. Some are unproven. One or two are hard to measure at all.
Each claim gets one of five ratings: Confirmed, Mostly True, Unproven, Misleading or Not Supported. Nothing here is financial advice, and every figure was current as of October 1, 2026.
“Want the backstory first? Read our breakdown of how SpaceX bought xAI and what it means for Grok.”
- How I checked these claims?
- Claim 1: A Fable or GPT-6 level model in 2 to 3 months
- Claim 2: "Pole position" in about 6 months
- Claim 3: "Only 3 years old" versus 6 and 10
- Claim 4: The "second derivative" is still strong
- Claim 5: 780,000 GPUs now, 1 million next week
- Claim 6: Compute is the key to the AI race
- Claim 7: Grok 4.7 will exceed all current models
- Claim 8: SpaceX's data gives it an engineering edge
- Claim 9: Extra intelligence is pointless past a point
- Claim 10: Grok 5 will be AGI
- Claim 11: SpaceX bought its way into AI coding
- Claim 12: The AI business will pay for itself
- The full scorecard
- What to watch next?
- FAQ
How I Checked These Claims?
Musk's own posts were the starting point, but they were never the final word. For model quality, I leaned on independent benchmark data instead of launch pages. For money, I used SpaceX's quarterly results and major financial outlets. For anything Musk said about the future, I marked it as a prediction instead of a fact.
One caution before we start. Benchmark scores depend on the version of the test and the way a model is run. Different write-ups quote different absolute numbers for the same models, so compare scores only within one version of one index.
Claim 1: SpaceXAI Will Have a Fable or GPT-6 Level Model in 2 to 3 Months
Musk's first post was the most specific. Counting from September 24, "two to three months" points to late November or December 2026.
The best independent yardstick today is the Artificial Analysis Intelligence Index, which blends ten benchmarks. On that test, Grok 4.7 scores 46 while Claude Fable 5.1 and GPT-6 sit at 53 each. That is a real gap, not a rounding error.
It gets wider in agentic coding. On Terminal-Bench 4.0, Grok 4.7 managed 26%, against 60% for GPT-6 Astra and 55% for Claude Fable 5.1.
Musk's own roadmap treats this as a climb. On September 14 he said Grok 4.8 would be a noticeable step up, and that Grok 4.9 would reach Astra or Fable-class performance later on. In the days after the Grok 4.7 launch, he also described SpaceXAI as third in agentic coding, behind Anthropic and OpenAI.
So the target is believable inside his own plan. It is not proven by anything that has shipped yet.
Verdict: Unproven. Check back in December.
Claim 2: SpaceXAI Will Reach "Pole Position" in About 6 Months
Here is the problem with chasing the top: the top keeps moving.
OpenAI shipped GPT-6 Astra on September 3, 2026 at $10 per million input tokens and $50 per million output tokens. Early access went to a small group of enterprise customers, because Astra was the first OpenAI model to hit the company's Critical cybersecurity level.
That was a few weeks before Musk's post. By the time SpaceXAI ships a Fable class model, its rivals will likely have shipped something newer.
There is also no single scoreboard for "the lead." A model can win on price, on coding, on agent tasks or on raw reasoning. SpaceXAI may well lead on cost today, and the leaders may lead on everything else.
Musk attached a condition to his own claim. He said it holds only if his "second derivative" stays strong, which we look at in Claim 4. Six months from September 24 lands around late March 2027.
“Curious how OpenAI's newest model reshaped the race? See our explainer on the GPT-6 Astra AGI claim.”
Verdict: Unproven. It is conditional, and the finish line is undefined.
Claim 3: "Our AI Efforts Are Only 3 Years Old, vs 6 and 10"
This one is short and mostly checks out.
Musk's post said SpaceXAI is only three years old compared with six years for Anthropic and ten for OpenAI. The math is close. Anthropic was founded in January 2021 by former OpenAI staff, which makes it nearly six years old today. OpenAI started in 2015, and Musk himself co-founded OpenAI in 2015 alongside Sam Altman, which puts it at about ten and a half years.
The age gap is real. What it means is less clear. A young lab with huge funding can hire experienced people and skip old mistakes. A young lab can also lack the research habits that took the others years to build.
The framing is fair. The conclusion Musk draws from it, that youth plus acceleration equals a coming lead, is still just a prediction.
“The rivalry behind these numbers is messy. Read our look at the Anthropic and OpenAI civil war.”
Verdict: Mostly True.
Claim 4: "If the Second Derivative Remains Strong"
This phrase sounds like jargon, but it is simple. Growth is good. Growth that keeps speeding up is better. The second derivative measures that speeding up.
SpaceX gave real evidence on the input side. Its AI computing power was about 1.4 gigawatts at the end of the second quarter, up roughly 250% from a year earlier. That is genuine acceleration in capacity.
The output side is less dramatic. Artificial Analysis reported that Grok 4.7 gained only 2 points on its Intelligence Index over Grok 4.6, while its coding agent score rose 9 points. So the picture is uneven: a small lift in general reasoning, a bigger lift in coding.
Release timing tells a similar story. Musk promised Grok 4.7 within about ten days of September 2. It arrived on September 21, later than promised.
Verdict: Mixed. Inputs are accelerating fast. Model quality is improving at a steadier pace.
Claim 5: SpaceXAI Has 780,000 GPUs Now and 1 Million Next Week
I like claims that come with numbers, because numbers can be checked.
According to a summary of Musk's reply, Colossus 1 has 150,000 H100s, 50,000 H200s and 30,000 GB200s. Colossus 2 has 110,000 GB200s and 440,000 GB300s. Add those up and you get 230,000 plus 550,000, which is 780,000. A further 220,000 GB300s would take the total to exactly 1 million, with another 220,000 planned for November.
The arithmetic holds. The Colossus 1 figure also lines up with outside reporting that the site holds more than 220,000 Nvidia graphics cards.
There are two catches. First, these counts come from Musk, with no independent audit. Second, the same report says Anthropic agreed in May to use Colossus 1 to serve Claude Pro and Max subscribers. That means roughly 230,000 of those GPUs are not training Grok at all.
“Compute is the new arms race. See the AI arms race warning signs we tracked.”
Verdict: Plausible, but self-reported. The math checks out, and a large slice of the hardware serves a rival.
Claim 6: Hardware and Computing Power Are the Key to AI
Musk's argument is that whoever brings the most compute online fastest wins. He has real credentials here. His team brought the first Colossus cluster online in just 122 days, and SpaceX is now renting capacity to Google and Anthropic in deals that could bring in around $26 billion a year. Those contracts include a 90 day cancellation clause, which lets SpaceX reclaim hardware if Grok needs it.
Industry researchers noticed the speed early. SemiAnalysis documented how Colossus 2 became the first gigawatt scale datacenter in the world and how fast the Memphis build moved.
But compute alone does not make a better model. According to Artificial Analysis data, Grok 4.7 used roughly 81,000 output tokens per task at its highest setting, versus about 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra. Grok 4.7 is a hungry model. More thinking per task is one way to buy a better score, and it is not the same as a smarter model.
Compute is necessary. The evidence says it is not sufficient.
Verdict: Partly True.
Claim 7: Grok 4.7 Will Exceed All Current Models
This is the cleanest test in the whole list, because Musk made the prediction before the model shipped.
On September 2, he said Grok 4.7 would beat every model available and noted that the model sits on a 2.1 trillion parameter base. He did add that Anthropic would probably release improved models soon.
By September 14 the tone had shifted. Musk said Grok 4.7 was roughly on par with Opus 5.0, behind the newer Opus 5.1, and "better in some ways, worse in others." He also blamed a training issue where the model quit hard tasks too early.
The independent numbers, covered under Claim 1, show Grok 4.7 behind Claude Fable 5.1 and GPT-6 on the Intelligence Index.
Fair credit belongs here too. Artificial Analysis said the result brings SpaceXAI into the top four AI labs, and its coding agent score improved sharply. That is a real step forward. It is just not what was promised.
“Musk's predictions have a long history. Browse our collection of Elon Musk superintelligence quotes.”
Verdict: Not Supported by independent benchmarks.
Claim 8: SpaceX's Training Data Gives Grok an Engineering Edge
In that same September 2 post, Musk argued that SpaceX's training data is unique enough that he would be shocked if any model beat Grok 4.7 at real world engineering.
It is a clever argument. Real engineering data from rockets and satellites is rare. If it improves the model, you might expect strong results on engineering and coding tests.
We do not have an independent engineering benchmark that proves it. What we do have is mixed. Grok 4.7 trails Claude Fable 5.1 on several tests, including CursorBench, GDPval and Terminal-Bench. On Terminal-Bench 4.0 it scored 26%, even slightly below a cheaper open weight model at 27%. Epoch AI has described that benchmark as flawed, so I would not lean on one test.
The way a model is run matters too. Three different Terminal-Bench figures circulated at launch: 38% in xAI's own model card, 33% in the Coding Agent Index and roughly 26% in the standardized test. Same model, three scores, because the setup differed.
The edge may exist. It has not been shown yet.
Verdict: Unproven.
Claim 9: Extra Intelligence Is Pointless Past a Certain Level
This is Musk's most interesting point. He argued that once a model far exceeds the intelligence a task needs, more intelligence is pointless for that class of work.
He has a point. Nobody needs a genius to reformat a spreadsheet or write a polite reply to a customer. For routine work, speed and price matter more than raw brainpower.
Price is where SpaceXAI shines. Grok 4.7 costs $2 per million input tokens and $6 per million output tokens. GPT-6 Astra costs $10 and $50. That is about five times cheaper on input and more than eight times cheaper on output.
The sticker price overstates the savings, though. As Claim 6 showed, Grok 4.7 burns more tokens per task, so the real gap per finished job is smaller.
It is also worth noticing what this argument is. It is a case for value, and a good one. It is not a case for leadership. A company can be the best deal in town without having the best model.
Verdict: Mostly True as a principle. It does not support the "lead AI" claim.
Claim 10: Grok 5 Will Be AGI
Asked on September 14 how close Grok 4.8 would get to artificial general intelligence, Musk replied "That will be Grok 5." He then said Grok 5 might be better than anything. Neither post came with a benchmark, a definition of AGI or a release date.
That matters, because AGI has no agreed test. Without a definition, there is nothing to verify. A claim that cannot be checked is not wrong, but it is not evidence either.
The timeline track record is mixed. Back in late 2025, Musk said Grok 5 was delayed into early 2026. As of today, I could not find a published release date.
There is also an odd contrast. That same day, Musk publicly backed a call from leaders at Anthropic and OpenAI to slow the pace of AI development. Pushing toward AGI while endorsing a slowdown is not a contradiction you can fact-check, but it is worth noting.
“Leaders are asking the industry to hit the brakes. Read about the AI CEOs slow down call and the Coxon resignation thread.”
Verdict: Unproven. It cannot be tested until AGI is defined.
Claim 11: SpaceX Can Buy Its Way Into AI Coding
Not every claim here is a prediction. This one is a fact, and it checks out.
SpaceX agreed to acquire Cursor for $60 billion in an all stock deal announced June 16. It came days after SpaceX's blockbuster IPO, following an April arrangement that gave SpaceX the right to buy the company or pay a large break-up fee.
The deal closed fast. Bloomberg reported that the acquisition became effective on August 14, about two months after the formal announcement.
The strategy is smart. Coding is where Anthropic and OpenAI are strongest, and Cursor gives SpaceXAI a product developers already use. The caveat is that buying a popular tool is different from building the best model underneath it.
“Big money is reshaping AI talent too. See how one hire led to Ruoming Pang's $200 million Meta deal.”
Verdict: Confirmed.
Claim 12: The AI Business Will Pay for Itself
This is the claim underneath all the others, the idea that a huge bet on AI will turn into a huge business.
The growth is real. In the second quarter, SpaceX's AI segment brought in $2.56 billion in revenue, up 247% from a year earlier. The company also reported $14.1 billion of contracted sales from cloud services agreements in its official quarterly results.
The costs are real too. That same AI segment lost $1.26 billion on an operating basis. Nearly $16 billion of capital spending went to the xAI business in the quarter, and the stock fell after the report. AI capex roughly doubled in one quarter, from $7.72 billion to $15.83 billion.
So the business is growing fast and spending even faster. Whether that becomes profit depends on pricing, demand and whether renting compute to rivals continues. That 90 day cancellation clause cuts both ways.
“For the bigger picture of this rivalry, read our take on the Google vs OpenAI AI war.”
Verdict: Mixed. Revenue is surging. Profit is unproven.
The Full Scorecard
| # | Claim | Verdict |
|---|---|---|
| 1 | Fable or GPT-6 level model in 2 to 3 months | Unproven |
| 2 | "Pole position" in about 6 months | Unproven |
| 3 | Only 3 years old vs 6 and 10 | Mostly True |
| 4 | Second derivative stays strong | Mixed |
| 5 | 780,000 GPUs now, 1 million next week | Plausible, self-reported |
| 6 | Compute is the key to AI | Partly True |
| 7 | Grok 4.7 exceeds all current models | Not Supported |
| 8 | SpaceX data gives an engineering edge | Unproven |
| 9 | Extra intelligence is pointless past a point | Mostly True |
| 10 | Grok 5 will be AGI | Unproven |
| 11 | SpaceX bought its way into AI coding | Confirmed |
| 12 | AI business will pay for itself | Mixed |
The pattern is clear. The facts about money, deals and hardware mostly hold up. The claims about future model quality are mostly predictions that cannot be checked yet. And the one claim that could be tested right away, that Grok 4.7 would beat everything, did not hold.
What to Watch Next?
You do not need to take anyone's word for it. These dates will settle most of the debate:
- Grok 4.8 and 4.9: Musk said 4.8 would finish training in mid September. Watch for a release and for independent scores.
- November 2026: The next 220,000 GB300 GPUs are due. If they arrive, the compute claim looks better.
- Late November to December 2026: The two to three month window closes. Does a Fable or GPT-6 class SpaceXAI model exist, as measured by independent tests?
- Late March 2027: The six month window closes. Check the Intelligence Index, coding agent rankings and price per task.
FAQ
Is Grok 4.7 better than GPT-6 Astra?
Not on the main independent index, where Grok 4.7 scores 46 against 53. It is far cheaper, though, and results vary by test.
Will SpaceXAI lead the AI race by March 2027?
Nobody can say for sure. It depends on how "lead" is measured, and on how fast OpenAI and Anthropic keep shipping.
Is Grok 5 really AGI?
It is a claim, not a result. There is no agreed definition of AGI and no published benchmark for Grok 5.
Where can I check the scores myself?
The Artificial Analysis Grok 4.7 page lists the independent results and updates as new models ship.
The Bottom Line
Musk's promise is bold, and some of the groundwork behind it is real: more compute, a major acquisition and fast revenue growth. But a promise is not a result. On the one test that could be run immediately, Grok 4.7 fell short of its billing.
The smartest move is to keep a scoreboard and revisit it in December and again next spring. If SpaceXAI delivers, the data will show it. If it does not, the data will show that too.
This article is for information only and is not financial advice.
Related Reading