Cursor and SpaceXAI Unveil Grok 4.5: Musk Says It Matches Opus 4.7 Performance

Hello, I’m Zhijian Jun. Cursor’s latest release is not just another "code‑writing" model.

The company’s new strategy, highlighted in the announcement, is to train the model on authentic development data, apply reinforcement learning on real engineering tasks, and integrate SpaceXAI’s compute and model stack.

Consequently, the launch of Grok 4.5 places Cursor at the center of today’s frontier model announcements, and the significance lies less in the name than in the collaboration.

According to Cursor’s official blog, the partnership with SpaceXAI yields the most intelligent model they have built so far and marks the first time the model’s target expands beyond software engineering.

Previously, Cursor’s self‑developed models—such as Composer 2.5 and earlier Composer series—were primarily optimized for coding scenarios. Grok 4.5, however, aims to cover software engineering, data science, finance, law, and broader knowledge‑work tasks, signalling a shift from a "code‑aware editor" to a component of a larger work system.

Elon Musk weighed in on X, stating that internal evaluations consider Grok 4.5 roughly comparable to Opus 4.7 in capability but significantly faster. He highlighted the combination of ability, speed, and lower cost as the model’s competitive edge.

Model competition has moved beyond "who is smarter"; developers now care about sustained runtime, error rates, and cost‑effectiveness. This is the core claim Grok 4.5 seeks to prove.

Benchmark results

Terminal‑Bench 2.1 : 83.3 % (near GPT‑5.5’s 83.4 %, above Opus 4.8’s 78.9 %).

SWE‑Bench Multilingual : 78.0 % (below Opus 4.8’s 84.4 %, slightly above GPT‑5.5’s 77.8 %).

DeepSWE 1.0 : 62.0 % (below Fable 5’s 66.1 % and GPT‑5.5’s 64.3 %, above Opus 4.8’s 55.8 %).

SWE‑Bench Pro : 64.7 % (below Fable 5’s 80.3 % and Opus 4.8’s 69.2 %, above GPT‑5.5’s 58.6 %).

These figures do not translate to an outright “top‑of‑the‑world” claim. Grok 4.5’s strengths are evident in terminal‑focused tasks, engineering execution, and cost efficiency. While it still trails behind Fable 5 and Opus 4.8 on tougher SWE benchmarks, its scores place it solidly within the leading tier.

Cursor adds a crucial disclaimer: scores for SWE‑Bench Pro and Terminal‑Bench are reported by third‑party models, the SWE‑Bench Multilingual score for GPT‑5.5 comes from Cursor’s internal runs, and CursorBench was omitted because an early snapshot of the Cursor codebase was unintentionally included in the training data, making its impact unclear.

This transparency, in the author’s view, enhances credibility. A serious model release should disclose which data are reliable and which require caution.

Pricing and availability

Grok 4.5 is bundled into Cursor’s subscription plans, offering generous usage limits with a doubled quota in the first week. The base model costs $2 per million input tokens and $6 per million output tokens; a faster variant is priced at $4 per million input tokens and $18 per million output tokens.

The model is currently accessible through Cursor’s desktop client, web interface, iOS app, command‑line interface, and SDK.