Grok 45 Delivers Opus Class Performance at Lower Cost But Questions Remain

The principal question surrounding Grok 4.5 concerns whether a model priced significantly below its competitors delivers correspondingly reduced performance. Based on available evidence, the pricing structure appears to miss the fundamental value proposition entirely.

SpaceX Corp released the model on Tuesday, marking its first launch since absorbing xAI and agreeing to acquire coding tool Cursor for $60 billion. Elon Musk characterised the offering as Opus-class, referencing Anthropic’s flagship family, whilst claiming superior speed, enhanced token efficiency and reduced costs.

The benchmark data presents a nuanced picture rather than an unequivocal victory. According to SpaceXAI’s published charts, Grok 4.5 surpasses Anthropic’s Opus 4.8 on two of four coding benchmarks whilst trailing on the remaining two. Anthropic’s Fable 5 leads most charts outright. Independent evaluation from Artificial Analysis positions Grok 4.5 fourth on its GDPval index for real-world agentic knowledge work, behind the latest Claude releases, with an Elo rating of 1543. The model therefore does not represent the most capable option available, and SpaceXAI’s own data makes no such claim.

Grok 4.5 carries a price of $2 per million input tokens and $6 per million output tokens, compared with $5 and $25 for Opus 4.8. The more significant metric, however, concerns consumption rather than unit pricing. On one software engineering benchmark, Grok 4.5 completed tasks using an average of approximately 15,900 output tokens against roughly 67,000 for Opus 4.8, representing a fourfold differential.

Tokens constitute the units of text a model processes and generates, forming the basis of customer charges. A model that charges less per token whilst consuming substantially fewer units delivers compounded savings. Artificial Analysis calculated the cost at $0.49 per completed task, describing the model as positioned clearly on the frontier for performance relative to cost. This addresses the value proposition directly: not half the capability at half the price, but marginally reduced performance at substantially lower total cost per completed task.

Grok 4.5 employed a distinct training methodology compared with most coding models. Rather than learning exclusively from static code, it absorbed genuine developer session data from Cursor, including debugging traces, multi-file changes and user corrections when the tool produced errors. This provides signal regarding how software actually undergoes repair, not merely how it appears when completed.

The model operates at approximately 80 tokens per second, supports a 500,000-token context window, and builds upon a 1.5 trillion-parameter foundation trained across tens of thousands of Nvidia chips. According to launch materials, its strengths cluster around long-running agentic tasks: constructing applications end-to-end from single prompts, working across multiple code repositories, and operating within Word, Excel and PowerPoint. The model also achieved top scores on a legal benchmark from Harvey, suggesting the training mix extends beyond engineering applications.

Vendor benchmarks remain vendor benchmarks, and independent testing remains limited. Grok 4.5 will remain unavailable in the European Union until mid-July, whilst the Cursor acquisition has not yet closed. Whether developers adopt the platform will depend upon how the model performs on their specific work, not on cost per task metrics presented in promotional materials.