SpaceXAI Launches Grok 4.5 - An 'Opus-Class' Model at a Fraction of the Cost, Trained Jointly with Cursor

SpaceXAI announced its new flagship model, Grok 4.5, on July 8, 2026¹. Positioned as the company’s strongest model for coding, agentic tasks, and knowledge work, it is also the first model trained jointly with the AI coding editor Cursor¹². It marks the company’s first model release since SpaceX absorbed xAI earlier this year and completed the largest IPO in history in June³⁴.

Elon Musk describes Grok 4.5 as “an Opus-class model, but faster, more token-efficient and lower cost”³. Pricing is set at $2 per million input tokens and $6 per million output tokens, and the model is available immediately in Grok Build, on all Cursor plans, and from the SpaceXAI API console¹.

How to Read the Benchmark Numbers

In the benchmark results SpaceXAI published, Grok 4.5 scored 62.0% on DeepSWE 1.0 (run within each provider’s recommended harness), which measures practical software engineering tasks, surpassing Claude Opus 4.8 (max, 55.75%). On Terminal Bench 2.1, which tests terminal operations, it scored 83.3%, nearly matching GPT 5.5 (xhigh, 83.4%)¹.

At the same time, even on SpaceXAI’s own charts, the top spot across all four coding benchmarks belongs to Anthropic’s Claude Fable 5 (max). On DeepSWE 1.1, measured on a harness run by third party DataCurve, Grok 4.5 scored 53%, trailing Fable (70%), GPT 5.5 (67%), and Opus 4.8 (59%)—results SpaceXAI also disclosed¹. Musk himself said the internal assessment is that Grok 4.5 is “roughly comparable to Opus 4.7, but much faster”³, so the realistic positioning appears to be “top-tier performance delivered with speed and price” rather than best-in-class capability. The data is also an interesting example of how rankings shift with measurement conditions—our guide to AI benchmarks covers how to read these numbers in detail.

Token Efficiency as a New Competitive Axis

What SpaceXAI emphasizes in this release is efficiency rather than absolute performance. Beyond fast serving at 80 TPS (tokens per second), the company claims “twice greater token efficiency than the latest leading models at the same tasks”¹. Specifically, it says Grok 4.5 resolves SWE Bench Pro tasks with an average of 15,954 output tokens—about 4.2 times fewer than Opus 4.8 (max) at 67,020¹.

Because output tokens translate directly into cost under API metered billing, the pitch is that the low unit price ($2 input / $6 output) and lower token consumption multiply together. As long-running agentic workloads become more common, the announcement reflects how model selection criteria are shifting from “quality of a single response” to “cost and time per completed task.”

What the Cursor Partnership Means

Grok 4.5 is a mixture-of-experts (MoE) model trained across tens of thousands of NVIDIA GB300 GPUs¹². Notably, training included trillions of tokens of Cursor data capturing a wide range of user interactions². SpaceX agreed on June 16 to acquire Cursor (maker Anysphere) in a $60 billion stock deal, so the development-side integration has taken shape ahead of the acquisition’s expected close in the third quarter⁴.

On Cursor’s side, Grok 4.5 is available immediately across desktop, web, iOS, CLI, and the SDK on all plans, with usage doubled for the first week². Cursor’s own model, Composer 2.5, remains on offer as a different model weight class². SpaceXAI is also offering free Grok 4.5 usage for a limited time in Grok Build and Cursor¹.

Beyond coding, plugins for Word, PowerPoint, and Excel are available on the Microsoft marketplace, supporting office work such as building multi-sheet Excel models and creating slides¹.

Availability and Caveats

Grok 4.5 is available from day one in Grok Build, on all Cursor plans, and from the SpaceXAI console (model ID: grok-4.5)¹. However, it is not yet available in the EU in any products or the API console, with availability expected in mid-July¹. Cursor also offers a fast variant ($4 input / $18 output) alongside the standard model ($2 input / $6 output)².

The coding agent market is heating up along multiple axes—performance, price, and trust—from geopolitical moves like Alibaba banning employee use of Claude Code to Claude Sonnet 5’s low-price strategy. With Grok 4.5 mounting a head-on price challenge—“Opus-class at less than a third of the unit price”—further price adjustments across the industry seem likely in an API market where switching costs are low.