Elon Musk has instructed Tesla employees to switch to Grok, the AI model developed by his xAI venture, now part of SpaceX. The move comes as Tesla imposes new spending limits on third-party AI tools, even as Musk acknowledges Grok lags behind its main competitors.
Tesla Shifts to Grok Amid New AI Spending Caps
According to a memo sent to staff, Elon Musk told Tesla employees to use Grok ‘when possible’ due to its lower token costs compared to other AI models. This directive follows Tesla’s decision to cap employee spending on third-party AI tools at $200 (~AUD 288) per week, a limit that applies to Anthropic, OpenAI, and Google models but exempts xAI’s Grok.
Musk also encouraged engineers to provide direct feedback on Grok by emailing him. Tesla has been internally testing beta versions of Grok for several months, with xAI product lead Andrew Milich collaborating with staff to address issues.
Grok 4.5 Performance and Benchmark Results
xAI released Grok 4.5 this week, describing it as ‘Opus-class.’ However, benchmark data shows Grok 4.5 ranks 9th overall on a multi-domain leaderboard with a score of 76.3, trailing behind several models from OpenAI, Anthropic, and Google. Its coding score is 68.6, the lowest among models on the board.
On the DeepSWE 1.1 coding benchmark, Grok 4.5 achieved 53%, compared to 70% for Anthropic’s Claude Fable 5. Additionally, Grok 4.5’s cost per task is about $0 (~AUD 0), significantly lower than Claude Fable 5’s $2 (~AUD 2) per task.
Internal Preferences and Model Limitations
Despite the push to adopt Grok, four people familiar with Tesla’s internal usage report that engineers generally prefer Anthropic’s Claude for daily development work. Musk himself acknowledged the performance gap, stating, ‘In fairness, Fable is definitely better than Grok 4.5, but most tasks don’t require Fable-level capability.’
A recent issue also emerged when Cursor, a coding startup involved in Grok’s development, disclosed that a snapshot of its own codebase was included in Grok 4.5’s training data. The affected metric was excluded from published comparisons, and the data will be removed from future models.
Competitor Comparison
Tesla engineers have access to several leading AI models, including Anthropic Claude Opus 4.8, OpenAI GPT-4, and Google’s AI offerings. Anthropic Claude Opus 4.8 is priced at $5 (~AUD 7) per million input tokens and $25 (~AUD 36) per million output tokens, and is widely used for development tasks. OpenAI’s GPT-4 and Google’s models are also available, though Tesla has set explicit spending limits for these tools.
Compared to these competitors, Grok 4.5 is less expensive but ranks lower in both general and coding benchmarks, which has influenced internal preferences among Tesla engineers.
Historical Context
Tesla’s decision to control AI tool spending comes during a period of declining market share and sales. The company delivered 1.79 million vehicles globally in 2024, marking its first annual decline in over a decade. U.S. market share dropped to approximately 46% in 2025, down from 49% in 2024, and year-over-year sales fell by 8.6%.
Why this matters
Tesla’s move to prioritize Grok highlights the company’s focus on reducing operational costs as it faces increased competition and slowing sales. While Grok offers lower usage costs, its performance lags behind leading AI models, raising questions about the impact on engineering productivity.
The decision also underscores the influence of leadership in technology choices, especially when those choices benefit ventures linked to the CEO.
What this means for buyers
For Tesla customers and investors, the shift to Grok may affect the pace and quality of software development at the company. Buyers should monitor how these internal changes influence Tesla’s product updates and innovation, especially as the company navigates a more competitive EV landscape.
Key Specs
- Grok 4.5 Multi-Domain Leaderboard Rank: 9th (score: 76.3)
- Grok 4.5 Coding Score: 68.6
- Claude Fable 5 Coding Score: 70% (DeepSWE 1.1 benchmark)
- Grok 4.5 Cost Per Task: $0 (~AUD 0)
- Claude Fable 5 Cost Per Task: $2 (~AUD 2)
- Tesla AI Tool Spending Cap: $200 (~AUD 288) per week (excludes Grok)
- Tesla Global Deliveries (2024): 1.79 million
- Tesla US EV Market Share (2025): 46%
Source
Images sourced from: notateslaapp.com, pixabay.com
