Elon Musk is pushing Tesla Inc. (TSLA) employees to abandon competing AI models in favor of Grok 4.5, the latest large language model from his artificial intelligence venture SpaceXAI, according to an internal memo obtained by The Information. The directive, which asks staff to use Grok “wherever possible,” underscores a growing push within Musk’s corporate empire to lock in captive demand for its own technology—even as independent benchmarks show the model trailing rival Anthropic’s Claude Fable 5 on critical coding tasks.
In the memo, Musk argued that Grok 4.5 offers significantly lower token costs than competing models, a financial edge he wants Tesla to exploit. The mandate arrived just days after Tesla imposed a strict $200 weekly spending cap per employee on external AI tools from Anthropic, OpenAI, and Google. Grok, notably, is exempt from the limit. The move effectively funnels engineers toward the in-house model by making rival products either too expensive or administratively off-limits.
The push comes as SpaceXAI released Grok 4.5 on Wednesday, billing it as an “Opus-class” model that is faster, more token-efficient, and cheaper than Anthropic’s Claude Opus 4.8. According to the launch material, Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, compared to $5 and $25 respectively for Opus 4.8. But the sharper number is token consumption: on one software engineering benchmark, Grok 4.5 used roughly 15,900 output tokens to complete a task, against about 67,000 for Opus 4.8. That translates into a cost per completed task of $0.49, according to Artificial Analysis, compared to far higher bills for rival models.
Despite the cost advantage, Grok 4.5’s raw performance is a mixed bag. SpaceXAI’s own charts show the model beating Opus 4.8 on two of four coding benchmarks and losing on the other two. Anthropic’s Fable 5 leads most of those charts outright. On the independent Artificial Analysis Intelligence Index, Grok 4.5 ranks fourth with an Elo rating of 1,543, behind Fable 5, GPT-5.5, and Opus 4.8. The model’s coding score on SWE-Bench Pro sits at 64.7%, well below Fable 5’s 80.4%.
Musk himself has not denied the gap. “Fair to say that Fable is better than Grok 4.5, but most tasks don’t need Fable-level capability,” he wrote on X. In a separate post, Musk went further, calling his earlier skepticism about Anthropic “clearly wrong” and acknowledging the company as “obviously currently the leader in AI.” He also pledged not to weaponize SpaceXAI’s compute access against the Claude maker, writing: “I would never cut them off in a way that hurt them badly, even as a competitor. That’s not my style.”
Despite Musk’s directive, Tesla engineers have shown a clear preference for Anthropic’s Claude in their daily development work, according to four people familiar with the matter. The internal testing of Grok has been ongoing for months, with SpaceXAI product lead Andrew Milich working alongside Tesla staff to iron out bugs. Engineers were also asked to email Musk directly with feedback on Grok’s performance.
The internal mandate is the latest example of Musk’s long-standing practice of tightly interconnecting his companies—Tesla, SpaceX, SpaceXAI, and The Boring Company—through talent sharing, technology transfers, and joint projects. Grok already powers in-car voice assistance in Tesla vehicles. Tesla earlier this year announced a $2 billion investment in xAI, which later transferred to SpaceX after the rocket manufacturer acquired the AI firm. xAI was subsequently renamed SpaceXAI.
Grok 4.5 was developed jointly with Cursor, the AI coding platform whose parent company SpaceX is acquiring in a deal valued at $60 billion. Cursor CEO Michael Truell posted that the model has “become the daily driver for many on our team,” though reviewers noted the endorsement comes from a party with a direct stake in the launch. The model was trained on real developer session data from Cursor—including debugging traces, multi-file changes, and user corrections—giving it a signal about how software actually gets fixed, not just how it looks when finished.
The broader AI landscape is shifting from a performance arms race to a price war. Meta Platforms Inc. (META) on Wednesday opened paid API access to its new model Muse Spark 1.1, pricing it at $1.25 per million input tokens and $4.25 for output—undercutting Anthropic’s high-end models. OpenAI’s GPT-5.6 series, released Thursday, offers tiered pricing from $1 to $5 per million input tokens. Chinese models from DeepSeek and others have been gaining share on platforms like OpenRouter, often priced 60% to 90% below U.S. frontier models, and now account for over 30% of weekly token consumption on the platform.
SiliconData’s LLM Token Expenditure Index has dropped roughly 22% since May, falling from $2.10 per million tokens to $1.64, as enterprises increasingly route simple tasks to cheap models and reserve premium ones for hard problems. The rise of AI agents—which consume far more tokens than simple chatbots—has made cost efficiency a boardroom priority. Gartner forecasts that the share of enterprise software with embedded AI agents will rise from under 5% last year to 40% by the end of 2026.
For Tesla, the internal push toward Grok is as much about corporate strategy as it is about cost. By steering thousands of engineers onto SpaceXAI’s infrastructure, Musk can generate real-world usage data and feedback at a scale most AI labs can only dream of—while simultaneously starving competitors of a marquee enterprise customer. Whether Tesla’s engineers actually switch will depend on how Grok 4.5 performs on their own workflows, not on cost per task in a vendor chart.