The obvious question about Grok 4.5 is whether a model priced at a fraction of its rivals delivers a fraction of the performance.
The answer, on the published evidence, is that the pricing is not the point.
SpaceXAI released the model on Tuesday, its first since absorbing xAI and agreeing to buy the coding tool Cursor for $60 billion.
Elon Musk described it as Opus-class, referring to Anthropic's flagship family, but faster, more token-efficient and lower cost.
Where it sits
The benchmarks tell a mixed story rather than a triumphant one.
On SpaceXAI's own published charts, Grok 4.5 beats Anthropic's Opus 4.8 on two of four coding benchmarks and loses on the other two.
Anthropic's Fable 5 leads most of those charts outright.
Independent evaluation from Artificial Analysis places 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.
So it is not the most capable model available, and SpaceXAI's own data does not claim otherwise.
Why the price is not the story
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, against $5 and $25 for Opus 4.8.
But the sharper number is consumption, not price.
On one software engineering benchmark, Grok 4.5 completed tasks using an average of about 15,900 output tokens against roughly 67,000 for Opus 4.8, a gap of more than four times.
Tokens are the units of text a model processes and generates, and they are what customers actually pay for.
A model that charges less per token and uses far fewer of them compounds the saving twice over.
Artificial Analysis put the cost at $0.49 per completed task and described the model as sitting clearly on the frontier for performance against cost.
That is the answer to the value question: not half the model for half the price, but a slightly weaker model at a substantially lower total cost per job done.
What makes it different
Grok 4.5 was trained differently from most coding models.
Rather than learning only from static code, it absorbed real developer session data from Cursor, including debugging traces, multi-file changes and the corrections users made when the tool got things wrong.
That gives it a signal about how software actually gets fixed, not merely how it looks when finished.
It runs at about 80 tokens per second, supports a 500,000-token context window, and is built on a 1.5 trillion-parameter foundation trained across tens of thousands of Nvidia chips.
Its strengths, per the launch material, cluster around long-running agentic tasks: building applications end to end from a single prompt, working across multiple code repositories, and operating inside Word, Excel and PowerPoint.
It also topped a legal benchmark from Harvey, suggesting the training mix reaches beyond engineering.
The catch
Vendor benchmarks are vendor benchmarks, and independent testing is still thin.
Grok 4.5 is unavailable in the European Union until mid-July, and the Cursor acquisition has not yet closed.
Whether developers switch will depend on how the model behaves on their own work, not on cost per task in a chart.
