Elon Rebuilt xAI From Scratch. Four Months Later, Grok Ties OpenAI.

Grok has been behind Anthropic and OpenAI all year. Every month of 2026, they shipped something better, and most of the time it wasn't even close.

On Wednesday, that ended. SpaceXAI released Grok 4.6. On the Artificial Analysis Intelligence Index, the third-party scoreboard the AI world actually watches, it scored a 61. That ties OpenAI's GPT-5.6 Sol. Only Anthropic's top two models are higher, at 63 and 62. Grok 4.5 scored 56 about five weeks ago. Grok 4.3 was 23 points below where this model is today.

If this sounds familiar, it's because in April I wrote that Elon Musk might win the AGI race. My argument was never that Grok was the best model. It was that he is the only one building everything himself, from the chips to the rockets, and that at some point raw compute wins. In July, I came back to that argument when Tesla's earnings showed what it was costing.

I'm not bringing this up to brag. I'm bringing it up because Wednesday is the first time it looks like my prediction might actually be correct.

He gutted the team, then bought a competitor

The comeback story is weirder than the score. Earlier this year, Musk basically gutted xAI. By the end of March, every one of the 11 co-founders besides Musk had left. He said publicly the company "was not built right the first time around" and he was rebuilding it "from the foundations up." SpaceX and Tesla executives came in to review the staff, and people got fired.

Then in June, days after SpaceX's record IPO, he agreed to buy Cursor, the AI coding tool, for $60 billion in stock.

Most companies that fire their founding team and drop $60 billion on a buyout are in for a rough year pulling it all together. Instead, about eight weeks after announcing the Cursor deal, SpaceXAI shipped a model that ties OpenAI's best on the main score, and it launched inside Cursor on day one. The deal is expected to close this quarter, and the model is already shipping in the product.

I want to be clear about something up front: I'm not going to sit here and tell you Grok is now the best model in the world. Anthropic still holds the top two spots on that score, and OpenAI's model still wins on some coding tests. But "right there with them" is a spot Grok hasn't been in all year. And the honest takeaway is simpler: Elon can basically just do anything he puts his mind to.

The price is the real move

Here's the number I can't stop thinking about. Grok 4.6 costs $2 per million input tokens and $6 per million output tokens. That's the same price as Alibaba's Qwen 3.8 Max. (Prompts over 200,000 tokens cost more, but the headline price is what everyone's quoting.)

Same level of smart, about a fifth of the output price. As someone who runs AI agents all day, cheaper tokens mean more work for the same money. We love to see it, because we all want AI to keep going down in price.

So why sell a top-tier model this cheap? Here's my read.

Anthropic and OpenAI live off token sales. That's basically their whole income. Elon doesn't need the token money. Starlink just did $4.3 billion in revenue in a single quarter and made a profit doing it. SpaceX reported sitting on $100 billion in cash and marketable securities. And when SpaceX reported earnings as a public company for the first time, its AI division did $2.56 billion in revenue in one quarter and lost $1.26 billion.

He might be selling these tokens at a loss right now. But I don't think the company is worried about it. If anyone can get his company to trust him enough to run a loss leader strategy, it's him. They probably figure this is the right market move at the moment.

And I wouldn't be surprised if that's where the exact numbers came from. Qwen 3.8 Max just got released, and Grok 4.6 costs exactly the same, $2 and $6. If SpaceX was already planning a loss leader anyway, they probably looked at the new Chinese model and thought, well damn, this thing is really good, we might as well match whatever they're charging right now.

I don't want to call Anthropic and OpenAI incumbents. These companies all came up around the same time. But in this fight, the roles are clear: two companies that have to protect their margins, and one guy who doesn't.

One more detail makes it even better. SpaceX is renting server space to Anthropic. The same company collecting Anthropic's rent is undercutting Anthropic's prices.

If the price war slows those two down even a little, it buys him time to leapfrog them. And even if this pricing is pure strategy, it's the smartest part of this release.

The company that could have made this move

The weirdest part of this whole story is Google.

Google could have made this exact move. More cash than anyone, its own chips, its own cloud. Instead, in early August, it moved Demis Hassabis out of the day-to-day leadership of its DeepMind AI lab. Jeff Dean, employee number 30 and one of the key people behind Gemini, left with three other senior researchers to start a new company. In June, one of Gemini's leaders left for OpenAI, and the Nobel Prize winner behind AlphaFold left for Anthropic. Google's next big Gemini model was supposed to ship in June. It still hasn't.

That's a lot of top talent walking out the door in one summer.

My theory, and it's just a theory: Google doesn't want to spend the money to win up front. Too many side quests going on inside that company pulling attention in different directions. And honestly? Maybe that's the right move. If the valuations in this space are as inflated as they look, the company that spends the hardest is the one with the most to lose.

In April, I watched the factory

One more thing that makes this even wilder. Musk says Grok 4.7 is targeting a release in about two weeks. He called it a 2.1 trillion parameter model, better than 4.6 in every way, except slightly slower to serve, with even better token efficiency. If that timeline holds, think about where that puts us. Grok 4.6 is already competing head-to-head with the two labs at the top of this field. If 4.7 really is better than 4.6 in every way, is that the model that leapfrogs them?

If that happens, that's crazy.

Everyone spent this year watching the scores, waiting for Grok to catch up. It caught up. But the part I'll be watching isn't the next leaderboard. It's whether the price holds.

Whoever wins this race, we all want AI to keep getting cheaper. This week, for the first time all year, the pressure on prices is coming from a model that's actually good.