Grok 4

// Overview

Grok 4 is xAI's flagship large language model, released July 9, 2025 as the company's first model positioned as a frontier reasoning system. It is the successor to Grok 3 and the foundation of the later Grok 4.1 and Grok 4.3 releases. Grok 4 was trained with roughly 10x more reinforcement-learning compute than Grok 3, using xAI's Colossus cluster of about 200,000 GPUs, and reinforcement learning was applied at pretraining scale to make tool use a first-class capability rather than an add-on.

What sets Grok 4 apart is native, RL-trained tool use: the model reasons while calling a code interpreter and searching the web and X (formerly Twitter) in real time, then folds the results back into its answer. It accepts text and image input and returns text. xAI shipped it in two forms — standard Grok 4 (a single reasoning agent) and Grok 4 Heavy, a multi-agent tier in which several copies of the model work a problem in parallel and reconcile their answers for higher accuracy on the hardest tasks.

At launch Grok 4 posted frontier scores across reasoning, math, and science benchmarks, and Grok 4 Heavy was the first system to clear 50% on Humanity's Last Exam (text-only subset). The standard model is served via the xAI API as grok-4-0709 with a 256K-token context window (128K in the Grok consumer app). It has since been superseded by cheaper, faster Grok 4.1 and 4.3 releases but remains accessible as a legacy model.

// Benchmarks

1. Humanity's Last Exam (with tools)41%
2. GPQA87.5%
3. AIME 202591.7%
4. HMMT 202590%
5. LiveCodeBench (Jan–May)79%
6. USAMO 202537.5%
7. ARC-AGI-215.9%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

xAI list price for the grok-4-0709 model. The standard Grok 4 tier; the Grok 4 Heavy multi-agent tier is offered through the SuperGrok Heavy consumer plan ($300/month).

Pricing source ↗

// Strengths

• Frontier reasoning, math, and graduate-level science performance at launch (GPQA 87.5%, AIME 2025 91.7%, HMMT 2025 90.0%)
• Native, reinforcement-learning-trained tool use — code execution plus real-time web and X search woven into its reasoning
• Strong abstract-reasoning generalization: first model to break the single-digit wall on ARC-AGI-2 (15.9%, independently verified by the ARC Prize Foundation)
• Grok 4 Heavy multi-agent tier pushes the hardest benchmarks higher (50.7% on Humanity's Last Exam, 100% on AIME 2025)
• Large 256K-token API context window for long documents and extended agentic sessions
• Vision (image) input alongside text

// Best for

• Hard STEM and competition-math problem solving (AIME/HMMT/USAMO-style reasoning)
• Graduate-level science Q&A and research assistance
• Agentic workflows that need a model to plan, run code, and search the web/X mid-reasoning
• Coding and code review aided by tool use and a large context window
• Long-document analysis using the 256K-token context window
• Research-grade tasks where Grok 4 Heavy's parallel multi-agent accuracy is worth the higher cost

// How to access

// Grok (flagship) — every version

The full lineage of the Grok (flagship) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.