xAI's Real Plan to Win Isn't Grok 4.6

0:00 This week, a model tied OpenAI's top score on the main public index the industry ranks them with. It runs at a fifth of the output price, and it shipped inside the code editor a lot of

0:10 engineers already keep open. The model is Grock 4.6, and it is the least interesting thing this company shipped this month. The day before, it shipped a product that no benchmark leaderboard

0:19 has a column for yet. An agent with its own computer that signs into your tools using your login and works while you sleep. Underneath both of those sits a third thing that neither launch post

0:30 mentions and that is where the money is. Two months ago, this company bought the editor outright. 3 months ago, it started renting its supercomput to a rival. And its founder told staff that

0:40 within 5 years, artificial intelligence will be 99% of what SpaceX is worth. So, the question worth the next 10 minutes is simple. If the model is that good, why is the model not the plan? We will

0:52 read the scoreboard properly, including two rows the launch post left off the chart. Then we follow the electricity because that is the layer that gets paid whichever model wins. Start with the

1:02 number that made the headlines. Artificial analysis runs a composite index built from nine separate evaluations and Grock 4.6 scores 61 on it. That is a tie with GPT 5.6 Saul, one

1:14 point behind Claude Fable 5, and two behind Claude Opus 5. Five points above where Grock 4.5 sat a month ago on the Agentic Evaluation, the one that measures paid knowledge work rather than

1:26 puzzles, it posts an ELO of 1753, second only to Opus 5. And it gets there in about 53 turns and half a billion input tokens, where Opus 5 takes 103 turns and four times the tokens. Then the price,

1:40 $2 per million tokens in, $6 out. Opus 5 charges 5 and 25. Soul charges 5 and 30. So on output, which is what an agent actually burns, grock is running at a quarter of anthropics rate and a fifth

1:54 of open AIS. That is the launch story and it holds up. But the same announcement carries two rows that do not flatter it. On Deep SEI, a software engineering evaluation, Gro scores 65.9

2:06 against Souls 73. Those are XAI's own published numbers, not a critics. The second row is Stranger, and it is the most useful thing in this video. Terminal Bench measures whether a model

2:17 can drive a real shell to finish a real job. On version 2.1, Grock scores 88.4, Soul scores 89.5. In practice, that is a tie. On version three of the same benchmark, Grock scores 26. Soul scores

2:31 34.6. Opus 5 scores 43.5. Same model, same week, same benchmark name, and the gap goes from one point to nearly 9. Nothing regressed. Version 3 rewrote the tasks

2:43 to push past the ceiling version 2.1 had hit. So those two numbers are not one measurement taken twice. They are two different exams. Any chart that puts them side by side is telling you a story

2:54 the data cannot support. There is a third row and it is not on the launch chart at all. On the omniscience evaluation, which tests whether a model will admit it does not know, Grock 4.6 6

3:05 answers correctly about 48% of the time and avoids inventing an answer about 66% of the time. Which means when it is out of its depth, it says so roughly two times in three and make something up the

3:16 other time. In a chat window that is survivable in an agent you have handed your login to. That third case is the one you plan around. Then there is the pricing detail the headline number

3:27 hides. The $2 rate applies to requests under 200,000 tokens. cross that line and the whole request bills at $4 in and 12 out, not the overflow. Every token in it, the window is 500,000 tokens. So the

3:42 moment you use more than 40% of the window this model is sold on, your bill doubles. The cash discount shrank too from 85% off to 75. Good model, not decisive. Which raises the question, a

3:54 leaderboard cannot answer. If models are converging, what are you actually buying? The answer arrived the day before the model did. On the 11th of August, XAI put Grockbot into early beta

4:05 and the pitch on its own page reads like a job description. AI teammates you can give real work to. They sign into your tools, use them just like you do, and come back with finished work. Every

4:16 account gets a persistent cloud computer, not an API key and not a bundle of integrations, but a machine with a browser, a file system, and a command line, which the bot drives the

4:26 way a person would, which means it works with software that has no integration to offer. And that is most software, the tool your finance team lives in, the admin panel, a vendor built years ago

4:37 and stopped touching, the dashboard with no export button. A conventional agent needs a connector for each of those. A bot with a browser needs a login. Then comes the part that turns a demo into a

4:48 product. You ask a bot to follow along once while you do a job. It saves that path as a routine and reruns it on a schedule or on demand. In the company's own words, "Show one a workflow today,

5:00 hand off the project by Friday." In the beta, a game developer handed a bot the job of replacing placeholder art across a codebase. It read the code, wrote a prompt for each asset, generated the

5:11 images, cropped them into transparent files, and wired them back in. 74 finished assets in about 2 hours. It is bundled, not sold. Grockbot comes with super Grohee with Cursor Ultra at $200 a

5:24 month and with Cursor Teams premium at 120 per seat. There is no free tier and no standalone plan, which tells you who it is aimed at. The subscription buys a weekly usage allowance whose size the

5:35 company does not publish. passed it. Work bills at the model's own token rates. And the documentation says there is no Grockbot specific spend cap yet. An agent that runs while you sleep bills

5:46 by the hour it runs, not by the head that hired it. Now, the detail that got past most of the coverage, and it is written in XAI's own documentation, not by a critic. All of your bots share one

5:57 cloud computer, one file stem, one set of browser sessions, one set of login available across your entire bot roster. The docs put it in a sentence worth reading twice. Do not use separate bots

6:08 as a security boundary. So the recruiting bot and the bot with production access are not two employees with two badges. They are two windows on one desk sharing one key ring. That is a

6:19 design choice documented upfront with least privilege guidance beside it. Connect only what a workflow needs. Start read only. Keep sending and purchasing and deleting behind an

6:30 approval. But an audit view of what the bots actually did is listed as coming, not shipped. So the trust model is the product and it is still being built. Which is why the corporate structure

6:40 matters more than it sounds. In February, SpaceX acquired XAI in an all stock deal, valuing SpaceX at a trillion dollars and XAI at 250 billion. XAI stopped being a separate company. The

6:54 models, the data centers, and the rockets landed on one balance sheet. In June, SpaceX went public on the NASDAQ. That same month, it agreed to buy Anyphere, the company behind the Cursor

7:04 editor, for $60 billion in stock. That is the largest acquisition of a venture-backed startup on record against roughly $4 billion in annualized revenue. 60 billion for an editor looks

7:16 absurd until you line it up with this week. Grock 4.6 is available in Cursor on day one. Grockbot is bundled into Cursor plans. So, what do you get when you buy the place? The work already

7:26 happens. You get the default before anyone compares models. But distribution is still only the second layer. The third is the one Musk spends his time on and it has nothing to do with models. In

7:37 May, Anthropic agreed to pay this company $1 and a4 billion a month for compute. Take that apart. Anthropic, the maker of the model sitting above Grock on the index, rents Colossus 1. The

7:49 Memphis cluster Grock was trained on. 220,000 graphics chips, 300 megawatt, running to May of 2029, worth more than $40 billion. The filing calls it monetizing unused compute capacity.

8:02 Google signed as well, $920 million a month for about $110,000 chips from October this year to June of 2029 as bridge capacity for its own agent platform. Those two contracts together

8:15 run at roughly $26 billion a year. Here is the arithmetic that makes the strategy visible. 1 and a4 billion a month is 15 billion a year. Divide that by 300 megawws and you get $50 per watt

8:28 per year. Now hold that number because Musk gave the same figure a different way. In the all hands, he said the value per watt is probably going to be $30 to $50 and that 10 gawatts online by the

8:40 end of next year would mean 300 to 500 billion a year in revenue. The top of his range is not a projection. It is a signed contract with a competitor and the numbers are already moving that way.

8:52 The company ended the quarter at 1.4 gawatt, expects to pass two by the end of this year and reported $2.6 billion of AI revenue for the quarter, up 213% from the quarter before. Musk told staff

9:06 AI revenue will pass every other space business around September. For scale, 10 gawatt running continuously is roughly what 8 million American homes use. That electricity has to come from somewhere.

9:18 And in South Haven, Mississippi, it came from 27 methane gas turbines that were running without an air permit, powering the second Colossus site. The NAACP, represented by the Southern

9:29 Environmental Law Center and Earth Justice, sued over exactly that, arguing those turbines are likely the largest industrial source of smog forming pollution in the 11count Memphis area.

9:40 That is a real cost paid by people who did not choose it and it belongs in the same story as the revenue. Here is the strongest case against everything I have said. If Grock were the plan, you would

9:50 not lease its training cluster to Anthropic for 3 years. A Moody's analyst put it plainly that the revenue does not strictly have to come from Grock. Read that way. Renting Colossus 1 looks like

10:01 surrender. And it is worth adding that the comparison chart in the launch post uses competitor figures the competitors published themselves. That is not one lab running four models through one

10:11 harness. It is a vendor scoreboard and the independent index is the only third-party number in the whole comparison. So the verdict for most developers, Grock 4.6 is the coding

10:21 model to default to right now. And the reason is price against parody, not the top of any chart. 61 against 63 at a fifth of the output cost is the knee of the curve and it is already inside the

10:33 editor you use. The exception is narrow and real. If your work looks like terminal bench version 3, long shell sessions where a wrong command costs you, pay for opus 5 instead. I would

10:44 still take grock at twice the price for long context knowledge work and I would not point a bot at a production admin account this month. But the model was never the plan. When one company sells

10:54 you the model, the editor it runs inside and the electricity your competitor rents by the megawatt, who is left to lose?