Grok 4.5 Review: xAI's Strongest Model — Free, Generates Games, Cursor Integration, 80 TPS

TL;DR

Grok 4.5 is xAI's most powerful AI model, released July 12, 2026. Trained on tens of thousands of NVIDIA GB300 GPUs, it hits #1 on SWE Marathon (29%), #3 on DeepSWE (62%), and delivers up to 80 tokens per second. It is currently free to use with any X (Twitter) account via Grok Build, and integrates directly with Cursor IDE. This review breaks down the benchmarks, coding capabilities, Agent workflow support, Office automation features, and whether Grok 4.5 deserves a spot in your AI toolkit.

Quick Facts

Why Grok 4.5 Matters

Grok 4.5 dropped on July 12 with a clear message: xAI is serious about competing at the frontier. Unlike previous Grok releases that felt like catching up, Grok 4.5 lands with benchmark scores that put it in the same conversation as Claude and GPT — and in some areas, ahead of them. Three things make this release significant:

1. Elite Coding Benchmarks — Not Just Hype

On DeepSWE 1.0, Grok 4.5 scored 62%, placing third globally behind only Claude Fable Max and GPT-5.5 XHigh. On SWE Marathon — a grueling test of long-form software engineering — it claimed the #1 spot at 29%, beating Claude Opus 4.8 Max and Claude Fable Max. On Terminal Bench 2.1 (83.3%) and SWE Bench Pro (64.7%), it also ranked #3. This is not a single lucky run — it is consistent performance across four independent coding benchmarks. For developers, this means Grok 4.5 is now a legitimate alternative to Claude and GPT for real engineering work.

2. Free Access — For Now

The biggest surprise: Grok 4.5 is currently free. You need an X (Twitter) account or a SuperGrok account — no separate subscription, no credit card. This is a limited-time offer, and xAI has not announced when it ends or what long-term pricing will look like. But right now, you can access a frontier-tier coding model for zero dollars. That changes the value equation for indie developers, students, and anyone who has been priced out of Claude Pro or ChatGPT Plus.

3. Agent-Native Design — Built for the CLI, Cursor, and Beyond

Grok 4.5 was trained with heavy emphasis on multi-step Agent workflows and reinforcement learning across hundreds of thousands of tasks. It supports Cursor IDE out of the box, has an official CLI, and powers Grok Build — a web app that handles not just coding but Excel modeling, PowerPoint slides, and Word documents. This is a model designed to do work, not just chat.

"Grok 4.5 is not here to be your chatbot. It is here to be your engineering assistant, your office analyst, and your Agent brain — and it is free to start."

Benchmark Deep Dive: Where Does Grok 4.5 Actually Excel?

DeepSWE 1.0 — Software Engineering Under Pressure

DeepSWE tests a model's ability to navigate unfamiliar codebases, identify bugs, make cross-file changes, and verify fixes. Grok 4.5's 62% score puts it ahead of most models and firmly in the top tier. The only models ahead are Claude Fable Max and GPT-5.5 XHigh — both of which cost real money. For context: at the time of writing, Grok 4.5 delivers this performance for free.

SWE Marathon — #1 in the World

SWE Marathon is a long-form software engineering test that demands sustained performance across extended tasks. Grok 4.5 scored 29% and took the #1 spot — ahead of Claude Opus 4.8 Max and Claude Fable Max. This is a significant achievement: it means Grok 4.5 is not just good at quick code snippets, but can handle marathon-level engineering workflows without losing coherence.

Other Key Results

• Terminal Bench 2.1: 83.3% — #3 globally, strong command-line proficiency

• SWE Bench Pro: 64.7% — #3 globally, solid professional-grade software engineering

• Token efficiency: ~2x improvement over previous Grok generation means faster, cheaper inference

Bottom line: Grok 4.5 is competitive with Claude and GPT on coding benchmarks, and outright leads on sustained engineering tasks. For a model that is currently free, the price-to-performance ratio is off the charts.

Training Infrastructure: The GB300 Advantage

xAI disclosed that Grok 4.5 was trained on tens of thousands of NVIDIA GB300 GPUs — the latest-generation hardware. The training process emphasized:

• Data deduplication and quality scoring — filtering noise before it reaches the model

• Domain-based data selection — prioritizing high-signal training examples

• High-coverage data mixing — ensuring broad capability without sacrificing depth

• Massive-scale reinforcement learning across hundreds of thousands of tasks

The RL phase specifically targeted multi-step software engineering, Agent workflows, technical reasoning, automated scoring, and model self-feedback. The training platform uses a highly asynchronous architecture, meaning Agents can run for hours continuously while tens of thousands of GPUs perform RL training in parallel. This architecture choice — prioritizing sustained Agent performance — is a different design philosophy from models optimized primarily for single-turn chat.

Grok Build: Code, Games, Office — All in One

Grok 4.5 is the default model powering Grok Build, xAI's web-based AI workspace. Beyond coding benchmarks, it delivers impressive real-world output — from full games to Office documents:

Real-World Demo: FPS Shooting Game

Grok 4.5 generated a complete first-person shooter game from a single prompt — working weapon system, targets, score tracking, and smooth browser gameplay. This isn't a tutorial follow-along; it's one-prompt generation of a fully playable game. For indie devs and rapid prototypers, this is a window into the future of game development.

Real-World Demo: Angry Birds Clone

Another community test: Grok 4.5 built a complete Angry Birds-style physics game — slingshot mechanics, destructible structures, pig enemies, and score tracking — all generated from a single natural-language description. The physics feel right, the art style is consistent, and the game is genuinely fun to play. This demonstrates Grok 4.5's ability to handle complex interactive logic, not just static page layouts.

PowerPoint, Excel, and Word

Beyond games, Grok Build handles the Office suite. It can construct complex Excel models from natural language (multi-sheet workbooks with formulas and annotations), use PowerPoint's native shape tools to generate flowcharts and slide decks, and write structured Word documents. A prompt like "Create a 5-slide Quarterly Business Review deck with charts" produces complete slides — layouts, data visualizations, and copy — ready to present.

Cursor Integration: Grok in Your IDE

xAI specifically trained Grok 4.5 alongside Cursor, the popular AI-powered code editor. The model has been optimized for real-world development workflows — not just generating code in isolation, but understanding project context, making cross-file changes, and working within existing codebases. If you use Cursor, Grok 4.5 is now available as an alternative backend alongside Claude and GPT models.

CLI Access: One Command to Start

For terminal-native developers, xAI provides a CLI that installs in one command. Windows users run irm https://x.ai/cli/install.ps1 | iex in PowerShell. Linux, macOS, and WSL users run curl -fsSL https://x.ai/cli/install.sh | bash. The CLI gives direct access to Grok 4.5 from the terminal — useful for scripting, automation, and Agent workflows that need headless operation.

Speed: 80 TPS Changes the UX

Grok 4.5 delivers up to 80 tokens per second, placing it in the "fast model" category. Token efficiency has roughly doubled from the previous Grok generation. In practice, this means:

• Near-instant responses for short queries

• Smooth streaming for long code generation

• Lower compute cost per query (which likely enables the free tier)

• Better UX for interactive Agent workflows where iteration speed matters

For comparison: many frontier models run at 20-40 TPS. Grok 4.5's 80 TPS makes it one of the fastest frontier-tier models available. If you value responsiveness — especially for coding where you iterate rapidly — this is a tangible advantage.

Competitive Landscape: Grok 4.5 vs. the Field

Grok 4.5 vs Claude (Fable Max / Opus 4.8 Max)

Claude still leads on DeepSWE (Fable Max ahead of Grok's 62%), but Grok beats Opus 4.8 Max on SWE Marathon. Claude has the larger ecosystem and more polished developer experience. But Claude costs money. Grok 4.5 is free. For a developer on a budget, that single fact outweighs a few benchmark points.

Grok 4.5 vs GPT-5.5 XHigh

GPT-5.5 XHigh edges ahead on DeepSWE but Grok 4.5 wins on SWE Marathon. OpenAI's ecosystem is more mature (function calling, Assistants API, fine-tuning). Grok's advantage is raw speed (80 TPS vs ~40 TPS) and current free pricing.

Grok 4.5 vs DeepSeek V4 Pro

DeepSeek wins on pricing at $2/$10 per million tokens (when the Grok free period ends) and has a mature open-weight ecosystem. But Grok 4.5 leads on SWE Marathon and Terminal Bench. For cost-sensitive production use, DeepSeek may win long-term. For peak coding performance today at zero cost, Grok takes it.

Grok 4.5 vs Kimi K3

Kimi K3 is open-source (weights released July 27), multimodal (text + image + video + 3D), and leads on Frontend Code Arena (#1). Grok 4.5 is closed-source, text-only, but faster (80 TPS vs ~40 TPS) and integrated into Cursor and Grok Build. If you need multimodal generation or self-hosting, go K3. If you want a fast, free, IDE-integrated coding agent, go Grok.

Who Should (and Shouldn't) Use Grok 4.5?

Use Grok 4.5 if you are:

• A software developer who wants a free, fast coding assistant integrated with Cursor

• An AI agent builder who needs a model optimized for multi-step workflows and sustained reasoning

• A data analyst or office worker who can use Grok Build for Excel, PowerPoint, and Word automation

• A student or indie hacker who cannot afford $20/month AI subscriptions

• A CLI-native developer who wants terminal-first AI access

Skip Grok 4.5 if you:

• Need multimodal generation (images, video, 3D) — go with Kimi K3, GPT-5.6, or Claude

• Want the absolute highest reasoning scores regardless of cost — Claude Fable Max or GPT-5.5 XHigh still lead

• Prefer open-source with self-hosting options — wait for Kimi K3 weights or use DeepSeek

• Don't have or don't want an X (Twitter) account — this is currently a hard requirement

• Are building production systems that need guaranteed long-term pricing — the free tier is temporary

The X Account Requirement: Friction or Feature?

The biggest barrier to Grok 4.5 is the X account requirement. Unlike ChatGPT or Claude, you cannot sign up with just an email address. This is by design — Grok is deeply integrated into the X ecosystem, and xAI is leveraging X's existing user base for distribution. For the 500+ million X users, this is zero friction. For everyone else, it means creating an account on a platform you may not want to use.

Whether this is a dealbreaker depends on your relationship with X. If you already have an account, you get a frontier AI model for free. If you don't, the sign-up friction is real — but given the current free pricing, it may still be worth the 2 minutes it takes to create an account.

FAQ

Is Grok 4.5 really free?

Yes — for now. xAI has made Grok 4.5 available for free with any X (Twitter) or SuperGrok account. This is a limited-time promotion, and long-term pricing has not been announced.

How do I access Grok 4.5?

Three ways: (1) Grok Build on the web with an X account, (2) Cursor IDE integration, (3) CLI installation via x.ai. All three are currently free.

Is Grok 4.5 better than Claude for coding?

On SWE Marathon, yes — Grok 4.5 is #1 globally. On DeepSWE, Claude Fable Max still leads (ahead of Grok's 62%). For most practical coding tasks, they are in the same tier, but Grok is free while Claude costs $20/month.

Can Grok 4.5 generate images or video?

No. Grok 4.5 is a text and code model. For image generation, use DALL-E, Midjourney, or Stable Diffusion. For video, use Kling, Runway, or Sora.

Will Grok 4.5 stay free?

Unlikely. The free tier is explicitly time-limited. xAI has not announced pricing yet, but expect a subscription model similar to ChatGPT Plus or Claude Pro when the promotion ends.

Does Grok 4.5 work in China?

X (Twitter) is blocked in China, so accessing Grok 4.5 requires a VPN or proxy. The CLI and API access may work without the X platform directly, but initial account creation still requires X.