Musk is pushing Tesla staff toward Grok 4.5 while conceding that Anthropic still has the stronger model. That is not a product victory. It is an expense policy doing the work of persuasion.
Elon Musk told Tesla employees to move more of their AI work onto Grok 4.5, the model built by SpaceXAI, and he didn't dress it up as the better tool. According to The Information, with Electrek later detailing the internal memo, Musk wrote that Anthropic's Fable was "definitely better than Grok 4.5," but said most tasks didn't need Fable-level capability. That is the move. If the cheaper tool is good enough, the expensive favorite gets pushed to the side.
The memo landed alongside a harder rule. Tesla imposed a $200 weekly cap on what employees can spend on third-party AI tools starting July 6, The Information reported, with Anthropic, OpenAI and Google all covered by the limit. Grok and Composer, xAI's coding product, were exempt. If you want to spend more on the outside tools, you have to ask a manager.
You don't need an MBA to see what that combination does. Cap the tools your engineers actually use. Exempt the tool you own. The choice starts to disappear before anyone has proved Grok is better.
The policy is doing more than cutting costs
Tesla's new cap would be less interesting if employees were already moving to Grok on their own. They weren't. Electrek reported, citing four people familiar with internal usage patterns, that Tesla staff largely favored Claude over Grok for day-to-day development work, even after xAI product leaders ran onboarding sessions and Musk sent a company-wide email pushing Composer.
That preference matters because engineers don't choose coding tools out of brand loyalty. They choose the thing that gets them through a pull request with fewer mistakes and less wasted time. Claude has earned that position inside a lot of developer workflows. Anthropic has had outages and usage-limit frustrations, but developers kept coming back because the model still handled complex coding work well enough to be worth the irritation.
Frankly, this is a pressure campaign. Musk runs Tesla, and SpaceX now houses xAI under the SpaceXAI structure. Tesla has also put money behind xAI, according to reporting cited by Electrek, so a spending limit on rival tools doesn't only trim software costs. It nudges Tesla's internal demand toward a related Musk company.
There is a defensible version of that argument. A company does not have to pay unlimited bills for every engineer's favorite subscription, especially when AI usage can run away quickly. The problem is the exception. Once Grok and Composer get a free pass while Claude, OpenAI and Google sit behind a weekly cap, this stops looking like simple budget discipline.
Cheaper is a real argument
Grok 4.5 launched publicly on July 9, and the cost case is the strongest part of Musk's pitch. Axios reported that the model is priced at $2 per million input tokens and $6 per million output tokens. That undercuts the most expensive frontier coding models. Musk has also argued on X that Grok can finish some tasks using far fewer tokens than Anthropic's Fable, making the gap larger on real work than the headline price suggests.
That is useful. If an internal code question or a documentation task can be handled at a fraction of the cost, you should not pay for the top shelf model every time. Most companies will eventually sort AI work this way: cheap models for ordinary work, stronger models for the moments where failure costs more than the token bill.
The coding case is messier. Investor's Business Daily reported that Grok 4.5 has narrowed the gap with OpenAI, Anthropic and Google, and Axios described it as aimed at coding and agentic workflows rather than a consumer chatbot. But the article's own framing still has Musk admitting Anthropic's model is stronger. That admission is the center of the story, not a stray line in the memo.
SpaceX's $60 billion deal for Cursor, confirmed in recent reports from AP and The Verge, gives Musk a real shot at improving that weakness. Cursor already sits inside the workflows of serious developers, and its data on how people actually code is more valuable than another slick demo video. But an acquisition is not the same thing as trust. Engineers care about what works on Tuesday afternoon, not what the combined company might ship next quarter.
You should notice the order here. Tesla did not wait for Grok 4.5 to win the room. It capped the room first.
That is the real test. If Grok 4.5 is cheap enough and competent enough, Tesla may save money without much damage. If it is not, the cost of the cap will show up somewhere less tidy than an expense report: slower debugging and engineers quietly routing the hardest jobs back to Claude whenever they can get approval.
This is not neutral procurement. It is Musk using one company he controls to create demand for another tool he controls, while acknowledging the rival tool remains better. Sometimes that works. Sometimes it just teaches your best technical staff that the spreadsheet won before the product did.
Also read: Meta pulls its Muse image AI tool from Instagram after three days of backlash over consent • OpenAI's safety chief exits as the company folds oversight into research • Meta pulled its AI image tool from Instagram three days after launch
