Elon Musk has directed Tesla employees to adopt Grok 4.5, xAI's latest large language model (LLM), even as he publicly concedes the tool ranks below competitors from Anthropic and OpenAI on performance benchmarks. The push comes alongside a new $200 weekly spending cap on third-party AI tools that pointedly exempts Grok, raising questions about whether cost efficiency or corporate self-interest is driving the decision.

On Friday, Musk sent a memo to Tesla staff asking them to switch to Grok "when possible," citing lower token costs, which measure how much text the model processes. He also invited engineers to email him directly with feedback on the model. The directive follows Tesla's announcement earlier in the week that it would cap employee spending on AI tools from Anthropic, OpenAI, and Google, but notably exclude xAI's Grok from that restriction.

How Does Grok 4.5 Actually Compare to Its Rivals?

The benchmark data reveals a significant performance gap. On a widely used multi-domain leaderboard, Grok 4.5 ranks 9th overall with a score of 76.3, trailing multiple models from OpenAI, Anthropic, and Google. More telling, its coding score of 68.6 is the lowest of any model on the board. On other benchmarks, the pattern repeats: Grok 4.5 barely reaches the bottom of the top tier on LiveBench, the same position held by open-source Chinese models that cost a fraction of the price.

On DeepSWE 1.1, a neutral coding benchmark that measures resolving real GitHub issues, Grok 4.5 scored 53% compared to 70% for Claude Fable 5, Anthropic's latest model. There was also a benchmark transparency issue at launch: Cursor, the coding startup SpaceX is acquiring for $60 billion, disclosed that an earlier snapshot of its own codebase was accidentally included in Grok 4.5's training data, inflating one of Grok's headline coding scores.

"In fairness, Fable is definitely better than Grok 4.5, but most tasks don't require Fable-level capability," Musk stated on X.

Elon Musk, CEO of Tesla and xAI

Musk's acknowledgment underscores the real trade-off: Grok 4.5 runs at roughly $0.13 per task on the leaderboard, versus $1.57 for Claude Fable 5. That's a significant cost difference, but it reflects a performance difference as well.

Why Are Tesla Engineers Resisting the Switch?

Despite months of internal testing and direct support from xAI product lead Andrew Milich, Tesla engineers have consistently preferred Anthropic's Claude for day-to-day development work, according to four people familiar with internal usage. The new spending cap and directive to use Grok appear designed to override that preference through financial pressure rather than technical merit.

The structure of the policy makes the intent clear: by capping spending on all competitors while exempting Grok, Tesla is not simply asking engineers to consider a cheaper option. It is steering them toward a specific tool because their CEO owns the company that makes it. This raises concerns about self-dealing, where corporate decisions prioritize one executive's financial interests over employee productivity or shareholder value.

Steps to Address Rising AI Tool Costs Without Sacrificing Performance

• Self-host open-source models: The open-weight AI field now delivers 90 to 95% of frontier capability at a fraction of the cost. Models like DeepSeek-V4 and GLM-5.2 can run on hardware Tesla already controls, eliminating per-token billing and moving costs to power and amortized compute only.
• Negotiate volume pricing: Rather than capping spending across the board, negotiate enterprise discounts with leading AI providers based on usage volume, which can reduce costs without forcing engineers onto inferior tools.
• Implement tiered tool access: Reserve expensive, high-performance models for tasks that genuinely require them, while using cheaper alternatives for routine work, allowing engineers to choose the right tool for each job.

The underlying goal of controlling AI spending is reasonable. Coding tools are expensive, and reining in costs makes business sense. However, the method matters. If cost is the primary concern, self-hosting open-source models would actually solve the spending problem without forcing engineers onto a demonstrably weaker product. That approach would lock in low, predictable costs without moving money from one Musk entity to another.

Grok 4.5 was released on Wednesday alongside Cursor, and Musk initially pitched it as "Opus-class," referring to Anthropic's high-end model. The benchmarks tell a different story, and Musk's own public comments confirm it. The question now is whether Tesla's engineers will accept the mandate, or whether the company's productivity will suffer as a result of using a tool they know is suboptimal for their work.