Grok 4.7 shipped Monday with coding and office-work gains, and Tesla turned on Grok Bot the next day. Musk now says his three-year-old AI operation could lead the field within six months.
Elon Musk has spent months reminding critics that SpaceX’s artificial intelligence effort is young. On Thursday, he put numbers on it. He posted that the company’s AI work is three years old, compared with six years for Anthropic and 10 for OpenAI, and predicted SpaceX will take the lead in about six months if its rate of improvement holds.
He had two fresh exhibits. SpaceXAI released Grok 4.7 on Monday, Sept. 21, pitching it as the company’s strongest model yet for coding and knowledge work. On Tuesday, Tesla switched on Grok Bot inside its vehicles, turning the in-car assistant from something that answers questions into something that carries out tasks.
How much of this week is a product and how much is a press cycle depends on which piece you look at.
SpaceXAI’s launch post says Grok 4.7 is built on a larger base model than Grok 4.6 and was trained with a longer reinforcement learning run weighted toward problems that take many hours to finish. The company says the model is better at verifying its own work and handling longer context.
The price did not move. Grok 4.7 costs the same as Grok 4.6: $2 per million input tokens and $6 per million output for prompts under 200,000 tokens, with cached input at 50 cents. At 200,000 tokens and above, the rates double to $4 and $12. The context window is 500,000 tokens.
The headline benchmark numbers come from the company, and the comparisons aren’t apples to apples. SpaceXAI reports 46.3% on CursorBench 4.0 and 71% on DeepSWE, but most of its launch table set Grok 4.7 at its top “xhigh” reasoning setting against Grok 4.6 at the lower “high” setting. DeepSWE was the exception, run at high for both.
Outside scoring is less flattering. On CursorBench’s chart of accuracy against cost, Grok 4.7 lands in the middle, still behind Anthropic’s Fable 5.1 at every price point. Musk himself had lowered expectations days before launch, writing that the model should come in roughly even with Claude Opus 5.0 rather than Anthropic’s newer release.
That leaves value as the sales pitch. SpaceXAI markets the model as running twice as fast at half the price of comparable models. That’s the sticker price, and the bill per finished task can look different. At xhigh, Grok 4.7 used about 81,000 output tokens per task in independent testing, compared with 36,000 for Grok 4.6 at high, and the same testing found it cost more per task than a pricier-per-token OpenAI model.
Musk’s Thursday post leaned on three points. His AI lab is young and still accelerating. Once a model far exceeds what a job requires, extra intelligence is wasted: “You don’t need … Newton-level intelligence in your toaster,” he wrote. And standing up computing power fast is the hard part, which he argues SpaceX has already proven it can do.
There’s a twist on that last point. In May, SpaceX agreed to give Anthropic the full capacity of its Colossus 1 data center, more than 300 megawatts. Musk is renting the same scarce hardware to one of the labs he says he will pass.
In a separate post, Musk said he is cautiously optimistic SpaceX will have a Fable- or GPT-6-level model in two to three months. He has a record of dates that slipped, including years of fully driverless Tesla forecasts.
The most concrete evidence came from inside Tesla. Yun-Ta Tsai, a senior staff engineer on Tesla’s AI team, said Grok Build teams spent the past few weeks tuning the harness for 4.7, and that many Full Self-Driving and Cybercab features he personally shipped were finished with help from his overnight agents. Musk highlighted the post Tuesday, saying Grok 4.7 is already doing real engineering work inside Tesla.
Neither man said which features were involved, and Musk gave no figures comparing the agents’ output with traditional engineering work. Nobody compared the results against Claude or OpenAI’s models either. It’s a real internal use case that remains unquantified. Tsai has said this before; in May he confirmed Grok Build had been used extensively in FSD and Cybercab development.
Tesla’s Tuesday post on X was blunt: “Grok in your Tesla can now do meaningful work for you.”
The rollout has two layers. Connectors hook Grok up to outside services such as Gmail and Google Calendar, and Grok Bot handles multi-step jobs by voice, such as placing a coffee order or checking out an Amazon cart. Early users have posted about placing pickup orders and ordering groceries from the car.
That’s a big step for an assistant that arrived in Teslas in mid-2025 as a fairly conventional voice assistant and only learned this summer to change climate controls and settings on command. This week’s change takes the assistant out of the car’s own menus and into the owner’s outside accounts. A dedicated “Chief of Staff” agent can write and send emails and add calendar entries, while shopping agents can check out using saved payment details.
There are limits. Grok Bot’s task features are restricted to SuperGrok Heavy subscribers, a tier widely reported this week at $300 a month, and Tesla still classifies the in-car assistant as a beta. It can place calls through a linked phone but can’t yet send texts. The feature runs on newer AMD-based vehicles with Premium Connectivity. Tesla says more subscription tiers will get access without giving a timeline, and it has said nothing about availability in Europe. The company also says it’s working on tighter integration between Grok and FSD for what it called a “chauffeur-like experience.”
A young lab shipped a better coding and office model this week, and a car company put an agent built on the same technology in front of drivers’ inboxes and payment cards. Engineers inside Tesla say the overnight agents are already writing code that ends up in the vehicle. The benchmarks still split, and the full agent sits behind a premium subscription.
The next few months should answer the harder questions. Tesla and SpaceXAI haven’t said which FSD or Cybercab changes came from those overnight runs. It’s also unclear how Grok Bot will hold up once thousands of drivers hand it their Gmail and saved cards. Musk has now put a date on his own prediction, so if the next Grok doesn’t close the gaps 4.7 left, readers will be able to check the six-month clock for themselves.
His bet is that low prices and a fleet of cars will matter more than leaderboard rank, because a model that’s already in the loop when the work ships is hard to replace.
Sources:
Elon Musk on X, Sept. 24, 2026 (three-point post)
Elon Musk on X, Sept. 24, 2026 (Fable/GPT-6 timeline)
SpaceXAI, “Introducing Grok 4.7”
BERI, Grok 4.7 cost per task at matched reasoning effort
VentureBeat, Grok 4.7 token consumption and ROI
LLM Stats, Grok 4.7 model page
LLM Stats, Grok 4.7 launch analysis
Yahoo Tech, xAI launches Grok 4.7
Axios, How Elon grew to love Anthropic
Forbes, Musk says SpaceX AI may surpass Anthropic and OpenAI in 6 months
TeslaNorth, Tesla engineers using Grok 4.7
Basenor, Grok is now working on Tesla self-driving
Not a Tesla App, Tesla adds Grok Bot
Dataconomy, Tesla launches Grok Bot in cars
Teslarati, Tesla integrates Grok Bot
Poc Network, Grok connected to inbox, calendar and files
AutoGuide, Tesla integrates Grok Bot
TechBriefly, Tesla launches Grok Bot
Drive Tesla Canada, Grok Bot hands-free tasks
TeslAnt, Grok Bot and Connectors launch
Loading comments...