Grok

Overview

Grok is an AI assistant developed by xAI (Elon Musk's AI company), deeply integrated with the X (formerly Twitter) platform. Grok's unique positioning is as the "**most outspoken AI**" — humorous, straightforward, unafraid of controversial topics, and one of the few AIs capable of directly analyzing real-time X tweets and trends.

**Major New Release: Grok 4.5 launched on 2026-07-09** — the first jointly trained model after SpaceX acquired AI coding company Cursor for $60 billion, based on a **1.5T parameter V9 base**, trained on trillions of real developer coding interaction data from Cursor. SWE-bench Pro 64.7% (surpassing GPT-5.5's 58.6%, slightly behind Opus 4.8's 69.2%), with token consumption for equivalent tasks only **1/4** of Opus 4.8 (15,954 vs 67,020 tokens). Musk himself positions it as "roughly equivalent to Opus 4.7, but much faster." The Cursor editor natively integrates Grok 4.5, and a larger 2 trillion parameter version will be released in August.

Previously, the flagship **Grok 4.3 (fully available on Amazon Bedrock as of 2026-06-18)** ranked first in multiple benchmarks, with the **lowest hallucination rate + 1 million token context + 2-10x intelligence per dollar compared to other frontier models**. Supporting product matrix: **Grok Build** (terminal coding Agent), **Grok-on-X**, **DeepSearch**, **SuperGrok / SuperGrok Heavy** subscription tiers.

Key Features

• Grok 4.5 + SpaceX×Cursor Joint Training (New): Released 2026-07-09, 1.5T parameter V9 base, trained on trillions of Cursor developer data, SWE-bench Pro 64.7%, token efficiency 4x that of Opus 4.8
• Grok 4.3 (Lowest Hallucination Rate + 1M Context): Fully available on Bedrock as of 2026-06-18, first in multiple benchmarks, lowest hallucination rate in the industry, 1 million token context
• 2-10x Intelligence per Dollar: Bedrock pricing $1.25/$2.50 per M tokens, unit intelligence 2-10x that of other frontier models
• X Platform Real-Time Data: Can analyze real-time tweets, trending topics, and social dynamics on X, one of the few AIs with social media real-time data capability
• No Excessive Censorship: Broader response range, more "outspoken" than ChatGPT and Claude, does not easily refuse controversial topics
• DeepSearch Deep Research: Multi-step deep search and research function, can break down complex problems for long-chain research
• Grok Build Terminal Agent: Parallel sub-agent efficiency improved 3x, headless mode can connect to CI for automated deployment, comparable to Claude Code
• Image Understanding and Generation: Supports image understanding and AI image generation (Flux / Aurora models)
• Bedrock Enterprise Integration: On AWS Bedrock alongside Claude / Llama / Mistral, with enterprise compliance, SLA, and private VPC all available

Use Cases

• Users who need to analyze X / Twitter social media dynamics
• Users who find other AIs overly censored and need an "outspoken" AI
• AWS / Bedrock enterprise users adding multi-model options
• Media, marketing, and research professionals needing real-time trend analysis
• Grok Build for terminal coding Agent and CI automation
• Tech enthusiasts interested in the Elon Musk / xAI ecosystem

Pros

• Grok 4.5 extremely token-efficient, cost for equivalent tasks only 1/4 of Opus 4.8
• Data + distribution synergy after SpaceX acquired Cursor, strong targeting for coding scenarios
• Grok 4.3 lowest hallucination rate in the industry, long context 1 million
• 2-10x intelligence per dollar on Bedrock compared to same-tier frontier models
• Exclusive X platform data, unique social real-time analysis capability
• Fewer response restrictions, more outspoken than ChatGPT / Claude
• X Premium+ subscription includes usage rights, friendly to fan ecosystem
• Complete supporting ecosystem with Grok Build / Grok-on-X / DeepSearch
• Extremely fast iteration pace, xAI is one of the fastest model upgrade vendors in 2025-2026

Pricing

**Grok 4.3 on Bedrock**: Input $1.25/M tokens, Output $2.50/M tokens (2-10x intelligence per dollar). **X Subscription Ecosystem**: X Premium $8/month (basic Grok); X Premium+ $40/month (Grok 4.3 + Grok Build + DeepSearch); **SuperGrok $30/month**; **SuperGrok Heavy $300/month** (Grok 4 Heavy + maximum parallel Agent scale). **xAI API**: Billed per token, same tier as Bedrock.

Summary

Grok completed its transformation from a "Twitter built-in toy" to a "frontier model competing with Claude / GPT" in 2026 — **the Grok 4.5 released on 7/9 is the first joint training result after SpaceX acquired Cursor**, focusing on coding/law/finance tasks, with crushing token efficiency (cost only 1/4 of Opus 4.8), and natively integrated into the Cursor editor. If you code with Cursor, Grok 4.5 is one of the most cost-effective model options currently; if you are an AWS user doing enterprise model selection, Grok 4.3 on Bedrock remains a strong option; if you do X content analysis or value the "outspoken" attribute, Grok is still irreplaceable. A larger 2 trillion parameter version will be released in August, worth continuous attention.

Version History

• Cursor 测试新型 AI 智能体集群：规划者+执行者分工，4小时通过80% SQL测试 (2026-07-20): Cursor tested a new type of AI agent cluster, which breaks down tasks into planners (using the most powerful model) and executors (using fast, low-cost models). When using Grok 4.5, the new cluster passed 80% of the SQL test suite within 4 hours, while the old cluster failed before the second hour. This system has been used to build browsers, fix vulnerabilities, and generate billions of tokens of synthetic training data.
• xAI 推出 Grok for Outlook 加载项 (2026-07-20): xAI today launched Grok for Outlook, a Microsoft 365 add-in that embeds the Grok agent into email inboxes for summarizing long email threads, drafting replies in the user's style, and organizing the inbox. The tool is available immediately to all paid X and SuperGrok users and can be added from the Microsoft Marketplace.
• 代理群（Agent Swarm）通过树状分解在构建 SQLite（Rust 版）任务中达到 80% 测试通过率 (2026-07-21): An experiment demonstrated that after decomposing tasks into a tree structure of planners and executors, the agent group achieved an 80% SQL test pass rate within four hours using Grok 4.5, while the old agent group failed before the second hour. The new system reached a peak submission speed of 1,000 times per second, prompting the team to build a dedicated version control system from scratch. This architecture has been validated in tasks such as building browsers, fixing bugs, and generating billions of tokens of synthetic data.
• xAI 发布 Grok CLI 并支持 /tutorial 命令 (2026-07-26): Download Grok Build and enter /tutorial http://X.ai/cli
• Grok CLI Release (2026-07-26): xAI releases Grok CLI and Grok Build toolchain, supporting /tutorial command
• Grok 4.5 + Cursor Joint Training (2026-07-09): First joint model after SpaceX acquisition of Cursor, 1.5T parameter V9 base, SWE-bench Pro 64.7%, token efficiency 4x that of Opus 4.8, integrated into Cursor editor
• Grok 4.3 + Bedrock Fully Available (2026-06-18): Lowest hallucination rate, 1 million context, fully available on Bedrock, 2-10x intelligence per dollar, $1.25/$2.50 per M tokens
• Grok 4 / Grok 4 Heavy + Grok Build (First half of 2026): xAI flagship upgrade + terminal coding Agent Grok Build launched, parallel sub-agent speedup 3x
• Grok 3 / Grok 3 Reasoning (2025): xAI's first reasoning-enhanced model, 'AI Grok Mode' went viral among X users
• Grok 1 / 2 (2023-2024): Launched after Elon Musk founded xAI, open-sourced 314B parameter MoE, laying the foundation