SpaceXAI and Cursor release Grok 4.5 at $2/$6 per million tokens

SpaceXAI has launched Grok 4.5. This marks the first joint model release since the $60 billion acquisition of Cursor. Trained on tens of thousands of GB300 GPUs in SpaceXAI's Memphis data centres, the model is positioned as a highly efficient engine for coding, agentic tasks and knowledge work. The launch represents a significant strategic move by SpaceXAI to establish itself as a serious competitor in the frontier model space.

The release lands in the middle of a significant week for frontier models. Grok 4.5 enters the market with aggressive pricing and a focus on token efficiency. It is now the default model in Grok Build and is available across all Cursor plans.

Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens. Cached input tokens are priced at $0.50 per million. The model is served at fast-model speeds of 80 tokens per second. This combination of pricing and performance creates a compelling value proposition.

The most notable metric is the cost per finished task. SpaceXAI reports that Grok 4.5 resolves tasks on SWE-Bench Pro using an average of 15,954 output tokens. This is approximately 4.2 times fewer output tokens than Opus 4.8 at its maximum setting, which requires 67,020 output tokens per task.

When multiplied across a month of agentic operations, this efficiency difference becomes substantial. A firm running 100 tasks per day would save approximately 5.1 million output tokens per day compared to Opus 4.8. At $6 per million tokens, this translates to savings of approximately $30 per day, or $600 per month.

Token efficiency is often overlooked in discussions of AI model capabilities. Raw benchmark scores capture peak performance, but cost per task determines practical affordability. A model that solves problems using half the tokens costs half as much to operate, regardless of the base price per token.

SpaceXAI has invested heavily in reinforcement learning to achieve this efficiency. The model was trained on hundreds of thousands of tasks, with automated and model-based grading. This focus on per-token intelligence means the model makes better decisions about which information to retain and which to discard.

Grok 4.5 is explicitly designed for more than just software engineering. It targets office work directly through plugins for Microsoft Word, PowerPoint and Excel. These plugins are available via the Microsoft Marketplace.

In Excel, the model can build complex multi-sheet models that incorporate web research. It can construct formulas that reference multiple sheets and even leave notes for future reference. In PowerPoint, it utilises native shapes to construct diagrams and design slide content intuitively.

For AEC firms purchasing AI capabilities by the token, the price-to-performance ratio is critical. The combination of high token efficiency and low base pricing alters the economics of running autonomous agents. A firm that previously could not afford to run complex agents continuously may now find it economically viable.

We observe that the integration into Microsoft Office applications provides an immediate entry point for project managers. The ability to generate complex spreadsheets and presentations directly within the native applications streamlines the reporting process. Project managers can delegate the assembly of status reports, cost forecasts and schedule analyses to the agent.

The model's reported strength in legal work is also relevant for AEC firms. Many project delivery tasks involve contract review, risk identification and compliance documentation. The ability to automate these tasks at scale could significantly reduce administrative overhead.

The joint development between SpaceXAI and Cursor demonstrates the value of owning both the compute infrastructure and the application layer. By training on trillions of tokens of real user interactions from Cursor, the model has been optimised for actual developer and agent behaviours.

That's fundamentally different from training on public internet data. The model learns from millions of real coding sessions, understanding not just what code is correct but how developers actually work. The result is better performance on real-world tasks.

It is worth noting that Grok 4.5 is not currently available in the European Union. Access for EU users is expected in mid-July. This limitation may affect adoption in European AEC firms.

Grok 4.5 is not positioned as the highest-accuracy model. On several coding benchmarks, Claude Fable 5 and GPT-5.5 achieve higher scores. However, the cost per finished task metric tells a different story. For firms running agentic workflows at scale, the efficiency advantage is more important than marginal benchmark improvements.

That reflects a broader shift in how AI models are evaluated. Peak performance on standardised benchmarks is less relevant than practical performance on real-world tasks at an acceptable cost.

• The aggressive pricing of $2/$6 per million tokens significantly lowers the barrier for deploying agentic workflows at scale, enabling smaller firms to afford continuous AI agent operation.

• Achieving tasks with up to 4.2 times fewer output tokens fundamentally changes the cost calculation for complex operations, with potential monthly savings of hundreds of dollars for active users.

• Direct integration into Microsoft Excel, PowerPoint and Word brings frontier AI capabilities into the primary tools used by project delivery professionals, eliminating tool-switching friction.

• The model's design focus on resolving long-running tasks makes it highly suitable for data-heavy AEC workflows, including complex spreadsheet assembly and presentation generation.

• Real-world training data from millions of Cursor sessions optimises the model for practical developer and agent behaviours, improving performance on actual tasks beyond what benchmark scores suggest.

To explore the implications of the shifting economics of AI deployment, subscribe to the Project Flux newsletter.

All content reflects our personal views and is not intended as professional advice or to represent any organisation.