Grok 4.5 in Cursor: Agentic Coding Just Got WAY Cheaper (Full Tutorial)

Study Guide

Grok 4.5 in Cursor: Building Cheaper, Faster Agentic Apps

In this roughly 13-minute tutorial, Leon van Zyl demonstrates how to use Grok 4.5 inside the Cursor IDE to build a polished interior-design website, complete with an interactive 3D room, animations, and AI-generated images, and then deploy it to production without ever leaving the editor. The central argument is cost: Grok 4.5 delivers flagship-level coding results at a fraction of the price of models like Opus 4.8 and GPT 5.5.

Why Grok 4.5 Matters: The Cost Story

Rather than rehash benchmarks, Leon focuses on the economics that actually change how you work:

  • Input pricing: Grok 4.5 is only $2 per 1 million tokens, versus $5 for both Opus 4.8 and GPT 5.5.
  • Output pricing: Grok 4.5 is $6 per 1 million tokens, versus $25 for Opus 4.8 and $30 for GPT 5.5.
  • Token efficiency: It uses roughly 4x fewer tokens to reach similar results as the flagship models.
  • Real-world cost: On comparable tasks, Leon found Grok about 17x cheaper than Opus 4.8 while still performing strongly on coding benchmarks.
  • Speed: The model is "blazingly fast," a product of the massive infrastructure it was trained on (a joint XAI and Cursor effort, reportedly trained on a very large GPU cluster backed by SpaceX and Tesla infrastructure).

Two Ways to Run Grok 4.5

There are two main harnesses for the model:

  • Grok CLI — a command-line tool from XAI, the native agent tool where Grok 4.5 is the default model.
  • Cursor — a clean graphical interface with an embedded browser so the agent can view and test the app itself. If you have used Claude Code or the ChatGPT desktop app, it will feel familiar. This tutorial uses Cursor.

Setup: Installing and Configuring Cursor

  1. Download Cursor from cursor.com/download and install it.
  2. Cursor is a paid product. The free/hobby tier gives some inference but does not include Grok 4.5, so you must upgrade to at least the individual plan to use Grok models.
  3. Sign in via the button in the bottom-left corner.
  4. The left sidebar shows your workspaces, letting you jump between projects and run multiple agents in parallel within a project.
  5. Prefer a traditional editor? Click Open IDE for a VS Code-style view. Toggle the chat panel with Ctrl + L.
  6. This tutorial uses the agent view, which makes building apps accessible even to non-developers.

Starting a Project and Selecting the Model

  1. Open a new project via the dropdown → Open Folder → create a new (empty) folder. All agent changes land in this folder.
  2. View folder contents with the button in the top-right corner; the file list starts empty.
  3. Change the model to Grok 4.5. Importantly, do not enable Auto mode or Max mode.
  4. Click Edit next to the model to set reasoning effort (low / medium / high). Leon chooses high.
  5. Optionally enable Fast mode to speed up inference, but note it costs significantly more.

Key Concept: Use Plan Mode to Survive Long Sessions

Leon gives a critical warning: many developers report that the agent struggles with long-running tasks and may crash, stall, or start making mistakes after about 30 minutes. His recommended safeguard:

  • Switch into Plan mode by holding Shift + Tab (pressing again cycles through other modes like debugging, multitasking, and ask mode).
  • In Plan mode the agent makes no code changes; it only discusses and produces an implementation plan, which gets stored in the project folder.
  • Because the plan is saved, if Grok crashes you can simply ask it to continue from where it left off.
  • Review the generated to-do list before executing, then approve the plan to build.

Key Concept: In-Editor Image Generation

A standout, rarely-discussed feature: Cursor gives its agents a tool to generate images (Leon suspects Nano Banana is used behind the scenes). This is useful for:

  • Generating stock images, avatars, and icons directly for your web app.
  • Bringing life to dull landing pages, just ask the agent to generate images.
  • Creating a mockup of the website beforehand, then using that mockup to drive the design system and look-and-feel.

Key Concept: Grok Is an Iterative Model (Mind the Context Window)

A small button shows context usage. Grok 4.5's total context window is only 256,000 tokens, versus roughly 1 million for Opus or GPT models. Because of this, Leon stresses that Grok is meant to be used iteratively:

  • Do not try to build one massive solution at once.
  • Start with a foundation (e.g., the landing page and design system), then add features incrementally in new sessions.
  • For extra pages, create new sessions. For a SaaS or full-stack app, add one feature at a time.

Iterating with Voice and Text

Once the first version is built, Leon opens the site: a hero section, a 3D scene, and generated images. He iterates rapidly, even using his voice, to fix the 3D scene being too zoomed in, add mouse-follow rotation, and remove a stray black wall. Because the model is so fast, each change takes only seconds, making iteration genuinely enjoyable.

Deploying to Production Without Leaving Cursor

The final section shows one-click deployment via an MCP integration (Hostinger in this example):

  1. Sign up for a hosting account and choose a plan/term.
  2. In the hosting dashboard, go to Developer Tools → API. It supports VS Code, Cursor, Claude Code, and Codex.
  3. Select Cursor, choose configure manually, and ensure Websites, domains, subscriptions/payments, and email marketing capabilities are selected.
  4. Copy the JSON config from the "copy JSON config" section.
  5. In Cursor, go to Settings → Tools and MCPs → Open Customize → New MCP Server, and paste the configuration.
  6. Return to the project and simply ask the agent: "Please can you deploy this project to Hostinger."
  7. On first run the agent asks you to authenticate your hosting account in the browser; click authenticate.
  8. The agent commits, pushes, and returns a live URL. Custom domains use the same process.

Practical Takeaways

  • Cost changes behavior: Grok 4.5's low price and token efficiency make heavy iteration affordable.
  • Always start in Plan mode and save the plan so you can recover from crashes on long tasks.
  • Respect the 256K context window — build iteratively across fresh sessions rather than in one giant run.
  • Use in-editor image generation for stock art, icons, and design mockups.
  • Deploy via MCP so you never touch the hosting dashboard again after setup.
  • Do not use Auto or Max mode; do choose high reasoning effort for quality.

Notable Quotes

"Everyone's kind of talking about Grok 4.5 and the benchmarks, but no one is showing you the proper way to use it."

"On similar tasks I've actually found that Grok is about 17 times cheaper than something like Opus 4.8, and it also performs exceptionally well on the coding benchmarks."

"Grok 4.5 is meant to be an iterative model. So basically you shouldn't build one massive solution at once. Start with the foundation and then iteratively start adding more and more features in new sessions."

"Because this model is so fast, it's actually a pleasure to work with. It only took a few seconds to make that change."

YouTube