Image: agentpedia.codes / Cursor Grok 4.5 Launches in Cursor and the API, Built on SpaceX's Colossus Cluster
SpaceXAI and Cursor jointly released Grok 4.5 today, a 1.5-trillion-parameter model trained on Cursor user coding sessions and available at $2/M input tokens — roughly a quarter of the price of Claude Opus.
SpaceXAI released Grok 4.5 on July 8, opening it to Cursor users across desktop, web, iOS, and CLI the same day. Public API access followed on July 9. The model runs at roughly 80 tokens per second and is priced at $2 per million input tokens and $6 per million output tokens for the standard tier, or $4/$18 for faster inference.
This is the first public model to come out of the SpaceX-Cursor integration, roughly six weeks after SpaceX announced its $60 billion all-stock acquisition of Anysphere in June 2026.
What it is
Grok 4.5 is built on xAI’s V9 foundation model, a mixture-of-experts architecture with 1.5 trillion parameters, trained on SpaceX’s Colossus cluster across 220,000-plus NVIDIA GPUs. SpaceXAI supplemented that base with trillions of tokens from real Cursor user coding sessions: debugging traces, multi-file diffs, and back-and-forth refactoring work.
Elon Musk described it as “an Opus-class model, faster, more token-efficient, and lower cost.” The pricing makes that comparison concrete. Claude Opus-tier models run $15 to $75 per million tokens depending on the variant. At $2/$6, Grok 4.5 is positioned well below that range.
Benchmark results are mixed
The numbers SpaceXAI and Cursor published are worth reading carefully. On Terminal-Bench 2.1, Grok 4.5 scored 83.3%, which nearly matches GPT-5.5 (83.4%) but trails Fable 5 (84.3%) and Sol Ultra mode (91.9%). On SWE-Bench Pro at high effort, it scored 64.7%, below Fable 5’s 80.3%.
On DeepSWE 1.0, Grok 4.5 led with 62.0% against Opus 4.8’s 55.75%. On DeepSWE 1.1, it dropped to 53%, behind Opus 4.8’s 59%.
Cursor acknowledged a notable issue in the launch post: the Cursor repository itself was accidentally included in the training data. That means in-IDE performance figures could be inflated, since the model may have effectively memorized parts of the codebase it was tested against. Cursor flagged the contamination openly, which limits how much weight to put on any Cursor-specific benchmarks.
One figure stands out despite the mixed results: on SWE-Bench Pro, Grok 4.5 used 15,954 output tokens on average versus 67,020 for Claude Opus 4.8. That’s a 4.2x efficiency gap that directly affects cost on long agentic tasks.
Who should use it
At $2/$6 per million tokens, Grok 4.5 is competitive for high-volume tasks where cost matters more than peak benchmark performance. If you’re running hundreds of coding sessions a day inside Cursor, the price difference against Opus-tier models adds up fast.
For single-session accuracy on hard engineering problems, the current benchmark data puts Fable 5 and Opus 4.8 ahead, especially on SWE-Bench Pro. The contamination caveat also means Cursor-specific evals shouldn’t drive the comparison.
The model is not yet available in the EU. Cursor’s Composer 2.5 remains available as a coding-specialized alternative for users who want to keep using that.
Sources: SpaceXAI — Introducing Grok 4.5, Cursor — Introducing Grok 4.5, agentpedia.codes — Grok 4.5: The SpaceXAI + Cursor Coding Model, Benchmarked, roo.beehiiv.com — Grok 4.5 Launched Today: What xAI’s Benchmarks Actually Show vs Opus 4.8