Grok 4.5: SpaceXAI's Leap with Cursor in the AI Race

Grok 4.5 just dropped, and the demo is the part worth talking about first.

SpaceXAI shipped Grok 4.5 on July 9, 2026, built on Cursor, and one of the showcase builds was a solar system simulator in Three.js — time scrubbing, realistic orbital motion — generated from a single prompt. That is a hard thing to ship from a prompt, and Grok 4.5 did it. That's the headline. Everything else is positioning.

Then the comparison. Musk claims Grok 4.5 matches Anthropic's Claude Opus 4.7 on capability, runs "much faster," and costs less: "The combination of capability, higher speed, and lower cost is what makes it competitive." That's a vendor claim from launch day, not an independent benchmark. Until third-party numbers land, it's a competitive pitch — not a settled fact.

What we can say from the launch coverage: Grok 4.5 is the first joint model from SpaceXAI and Cursor, weeks after SpaceX agreed to acquire Cursor at a $60B valuation. It's now the default model in Grok Build, and SpaceXAI says it's built to "handle complex, long-running tasks" — software engineering first, then legal and financial work. The demos went past code: complex Excel models with web search and multi-sheet formulas, PowerPoint diagrams from a prompt, notes for future reference.

That's a real launch. Let me walk through what's actually new, what the Grok 4.5 vs Claude Opus 4.7 comparison can and can't say, and where this lands in a stack you probably already have.

What Grok 4.5 actually is

Three concrete things, not the marketing copy:

1. A joint SpaceXAI + Cursor model. Cursor was the development substrate; the acquisition closed the loop — SpaceX now owns the IDE its model was trained against.
2. Default model in Grok Build. Whatever was there before is now Grok 4.5 unless you switch it.
3. Positioned for long-running, multi-step work. Software engineering is the lead use case, with legal and financial as the named verticals.

That's the launch in three lines. The rest of the announcement is the "what can it do" demo reel.

The Grok 4.5 vs Claude Opus 4.7 comparison, honestly framed

The source gives us exactly one data point on the head-to-head: Musk says Grok 4.5 "can match Anthropic's Claude Opus 4.7 model but is 'much faster,'" with lower cost on top.

That's three claims — parity, speed, cost. None has an independent number attached.

This is the part to be careful with. Vendor speed claims on launch day are exactly the kind of thing that gets walked back once independent benchmarks run. If you're making a tooling decision on this, hold the headline loosely until SWE-bench, Terminal-Bench, or whoever you trust publishes a number.

What we can say with more confidence: the positioning has changed. SpaceXAI used to be the lab that lagged Anthropic and OpenAI on programming and enterprise work. Now it's claiming parity with Anthropic on the model tier that matters most for that work. Whether the parity is real is the open question — and it's the question worth watching.

What the demos showed

The most concrete signal in the launch was the build-it-from-a-prompt reel. SpaceXAI shipped these as worked examples:

• A Three.js solar system with time controls and "realistic movement." Not a static rendering — time scrubbing, orbital motion.
• Complex Excel models generated using web search and multi-sheet formulas, with notes for future reference.
• PowerPoint diagrams from a prompt.
• Drafting (the announcement cut off here).

The Excel + web search + multi-sheet combination is the one to watch. Most "AI for spreadsheets" stops at "fill this cell." Web search + multi-sheet + notes is a different shape — closer to a junior analyst with a browser than a formula completer.

The Three.js demo is also doing real work. A solar system simulator isn't just mesh + sphere; you need time-driven updates, camera controls, and a frame loop. If Grok 4.5 generates that from one prompt reliably, it's because Cursor-the-IDE trained against a lot of Three.js code, and that training data is now baked into the model. That's the use the SpaceX-Cursor acquisition bought.

How to actually use Grok 4.5 today

Two real paths as of launch.

Path 1: Grok Build (consumer surface)

Grok 4.5 is the default model. If you have a Grok account, you're already pointing at it.

# Open Grok Build
# Model selector shows:  Grok 4.5  ← default, no action required
#                       Grok 4   ← legacy, switch manually if needed

Path 2: Cursor (developer surface)

Cursor was the build substrate for Grok 4.5, and SpaceX now owns Cursor. The integration is the model itself. If you're a Cursor user, check the model selector after updating — Grok 4.5 should appear alongside Anthropic and OpenAI options.

// Cursor → Settings → Models
// Add model:  Grok 4.5
// Surface:    agent mode, composer, inline edits
//
// The prompt loop is identical to using Claude or GPT-4
// in Cursor — only the model id changes.

For the prompt-style demos (solar system, Excel, PowerPoint), the model is doing the work; Cursor is the surface. Same agent loop you're used to, different model behind it.

What the launch doesn't publish: an OpenRouter id, a direct API endpoint, or token pricing. SpaceXAI hasn't released those publicly. If you need raw API access, wait for the developer docs or route through Grok Build's export.

Where this lands in a stack you already have

This is the part that matters for builders, and the part the launch doesn't talk about.

AI coding models change every quarter. Grok 4.5 today. Claude Opus 4.7 yesterday. Whatever ships next quarter. The model is the volatile layer — it's the part you'll swap.

The part that doesn't swap is what the model writes into. Your component library, your design tokens, your cross-platform parity, the repo conventions that tell any agent where to put a new screen — those are the durable layer. When Grok 4.5 generates a Three.js demo from a prompt, it works because Three.js is well-documented and the patterns are stable. When it generates a screen in your actual product, it works because you defined the components it reaches for.

This is why the layer underneath matters: a model that "matches Claude Opus 4.7" is exciting today and unremarkable in six months. The repo conventions, the shared component API, the one place where web and mobile speak the same language — those compound. Every model that lands on top gets more productive, not less.

OTF sits in that layer. The same component renders on web, iOS, and Android from one API; the conventions are explicit; whichever agent you pick this month — Grok 4.5, Claude Opus 4.7, the next one — has a stable target to aim at. Use Grok 4.5. Use Claude. Use whatever ships next. The volatile layer stays volatile. The durable layer doesn't.

What to watch

Three things, in order of how much they'd change the picture:

1. Independent benchmarks on Grok 4.5 vs Claude Opus 4.7. SWE-bench Verified and Terminal-Bench are the usual suspects. Until they run, "much faster" is a vendor line.
2. API pricing and access surface. Launch-day parity claims don't mean much without a dollar number. The source gives us nothing on token pricing or a public API id.
3. Whether the SpaceX-Cursor integration tightens or fragments. The model and the IDE are now the same company; whether that shows up as Cursor-only features or as a clean public API is the open question.

Grok 4.5 is a real launch, the Cursor acquisition is a real bet, and the demos are real demos. Treat the Claude Opus 4.7 comparison as a vendor claim until the benchmarks land, and put the model in its right place — the volatile top of the stack, not the foundation.