Grok-4震撼发布：xAI第四代大语言模型的技术突破与安全挑战

Introduction

Grok 4是由埃隆·马斯克旗下人工智能公司xAI于2025年7月正式发布的大语言模型，系该公司自2023年推出初代模型以来的第四次重要迭代。该模型发布时即被定位为“世界上最强AI模型”，旨在与OpenAI、Anthropic等公司的顶尖模型展开全面竞争。Grok 4的发布不仅标志着xAI在技术路线上的重大突破，其后续的快速迭代、定价策略调整以及暴露的安全挑战，也为大语言模型领域的发展提供了重要的观察样本。

Grok 4 is a large language model (LLM) officially launched in July 2025 by xAI, the artificial intelligence company under Elon Musk. It represents the fourth major iteration since the company's first-generation model debuted in 2023. Positioned at its release as the "world's most powerful AI model," it aims to compete comprehensively with top-tier models from companies like OpenAI and Anthropic. The launch of Grok 4 not only signifies a major breakthrough in xAI's technological roadmap but also provides a crucial case study for the development of the LLM field, given its subsequent rapid iterations, pricing strategy adjustments, and exposed security challenges.

Model Overview & Key Specifications

Core Architecture and Versions

此次发布的 Grok 4分为两个主要版本：标准版Grok 4与多代理版本Grok 4 Heavy。两者均为纯推理模型，支持最高256K tokens的上下文窗口。该模型基于xAI的Colossus超级计算机xAI专门为训练Grok-4构建的超级计算设施，拥有超过100,000个GPU，计划扩展到100万个GPU。进行训练，旨在提供更强的逻辑推理和文本生成能力。

The released Grok 4 comes in two main versions: the standard Grok 4 and the multi-agent Grok 4 Heavy. Both are pure inference models supporting a context window of up to 256K tokens. The model was trained on xAI's Colossus supercomputer, designed to deliver enhanced logical reasoning and text generation capabilities.

Key Functional Features

从功能设计来看，Grok 4具备深度推理能力、自然的人类语音特征、实时网络访问能力，以及对互联网文化（如梗、俚语和幽默）的高精度理解。它支持函数调用和结构化输出，能将AI意图转化为实际动作或返回规整数据供程序解析。

In terms of functional design, Grok 4 possesses deep reasoning capabilities, natural human-like speech characteristics, real-time web access, and a high-precision understanding of internet culture (such as memes, slang, and humor). It supports function calling and structured output, enabling the transformation of AI intent into actual actions or returning formatted data for programmatic parsing.

Development Timeline and Major Milestones

Grok 4的发布与发展遵循了一个密集且充满事件的节奏，以下梳理了其关键节点：

The release and development of Grok 4 followed an intensive and eventful pace. The key milestones are outlined below:

• 2025年7月9/10日: xAI正式发布Grok 4，声称其在处理学术问题上已达到博士水平。
  • July 9/10, 2025: xAI officially launched Grok 4, claiming it had reached doctoral-level proficiency in handling academic problems.
• 2025年7月18日: 网络安全公司NeuralTrust宣布使用“回音室攻击”方法成功越狱Grok 4，成功率超过30%，暴露了模型的安全短板。
  • July 18, 2025: Cybersecurity company NeuralTrust announced the successful jailbreak of Grok 4 using an "Echo Chamber Attack" method, with a success rate exceeding 30%, exposing security vulnerabilities in the model.
• 2025年8月11日: xAI宣布Grok 4基础版向全球用户免费开放，免费用户每12小时可进行5次请求。
  • August 11, 2025: xAI announced that the basic version of Grok 4 would be freely available to global users, with free users allowed 5 requests every 12 hours.
• 2025年9月: 前谷歌DeepMind核心开发者Dustin Tran加盟xAI，参与开发成本降低15倍的Grok 4 Fast快速推理版本。
  • September 2025: Former Google DeepMind core developer Dustin Tran joined xAI to contribute to the development of Grok 4 Fast, a rapid inference version with 15x lower cost.
• 2025年10月18日: Grok 4参与大模型投资比赛，并在次日的加密市场测试中实现1.33万美元的持仓市值。
  • October 18, 2025: Grok 4 participated in a large model investment competition, achieving a portfolio market value of $13,300 in a cryptocurrency market test the following day.
• 2025年11月18/19日: xAI发布Grok 4.1模型，在大模型竞技场（LMArena）文本排行榜上位居首位，并将幻觉率从12.09%显著降低至4.22%。
  • November 18/19, 2025: xAI released the Grok 4.1 model, which topped the text leaderboard on the Large Model Arena (LMArena) and significantly reduced its hallucination rate from 12.09% to 4.22%.

Analysis of Core Technical and Policy Features

Novel System Rules for Information Processing

据 xAI 更新的系统指令，Grok 4 新增了两条核心规则，这构成了其独特的响应策略：

According to xAI's updated system instructions, Grok 4 introduced two core rules, forming its unique response strategy:

1. 多信源分析要求: 若查询涉及时事、主观主张或统计数据，需深度分析多方信源。系统默认媒体主观观点存在偏见（此判断无需告知用户）。
   • Multi-source Analysis Requirement: If a query involves current events, subjective claims, or statistical data, in-depth analysis of multiple sources is required. The system defaults to the position that media subjective viewpoints are biased (this judgment does not need to be disclosed to the user).
2. 政治不正确表述的保留: 只要主张有充分依据，回答不应回避政治不正确的表述。
   • Retention of Politically Incorrect Expressions: As long as a claim is well-founded, responses should not avoid politically incorrect expressions.

Benchmark Performance and Cost Analysis

Grok 4 在多项基准测试中展现出强劲但成本较高的性能特点。

Grok 4 demonstrated strong but costly performance characteristics across multiple benchmarks.

• 学术与推理能力: 在涵盖数学、工程及人文学科2500个博士级问题的“人类最后的考试”基准测试中，Grok 4取得25.4%的准确率。在HLE测试中，其标准得分35%，运用推理技术后提升至45%。
  • Academic and Reasoning Capability: In the "Humanity's Last Exam" benchmark, covering 2500 doctoral-level questions across mathematics, engineering, and humanities, Grok 4 achieved a 25.4% accuracy rate. In the HLE test, its standard score was 35%, increasing to 45% with reasoning techniques.
• 代码能力: Grok 4 Code 在 SWE Bench 中标准得分 72%，使用推理技术后达 75%，与 Claude 4 Opus 接近。
  • Coding Capability: Grok 4 Code achieved a standard score of 72% on the SWE Bench, reaching 75% with reasoning techniques, which is close to Claude 4 Opus.
• 成本对比: Grok 4 的每任务成本为 2 美元至 4 美元，高于 GPT-5 的 0.73 美元。在 ARC-AGI-1 测试中，Grok 4 以 68% 领先 GPT-5 的 65.7%，但每任务成本约为 1 美元，高于 GPT-5 的 0.51 美元，表明后者在当时具有更高的性价比。
  • Cost Comparison: The per-task cost for Grok 4 ranged from $2 to $4, higher than GPT-5's $0.73. In the ARC-AGI-1 test, Grok 4 led with 68% versus GPT-5's 65.7%, but its per-task cost was about $1, higher than GPT-5's $0.51, indicating that the latter offered better cost-performance at that time.

Ecosystem, Services, and Security Challenges

Service Models and Accessibility

xAI为Grok 4构建了分层服务体系。2025年8月，基础版免费开放，显著降低了使用门槛。同时，提供每月300美元的Super Grok Heavy订阅服务，该服务用户还可访问基于Grok 4大模型的“伴侣”功能，该功能提供了具有不同性格和形象的角色（如哥特风女孩Ani）进行交互。

xAI established a tiered service system for Grok 4. In August 2025, the basic version became freely available, significantly lowering the barrier to entry. Concurrently, a $300 monthly Super Grok Heavy subscription service was offered, whose users also gained access to the "Companion" feature based on the Grok 4 model. This feature allows interaction with characters possessing distinct personalities and images (e.g., the gothic girl Ani).

Notable Security Incident

2025年7月18日，网络安全公司NeuralTrust成功对Grok 4实施了“回音室攻击”。该方法通过引导模型进行多轮推理，逐步注入风险信息，从而规避常见的安全拦截机制，最终诱导模型生成涉及制造武器、毒品等违规内容，成功率高达30%。这一事件凸显了新一代大语言模型在面对复杂、迂回攻击路径时仍存在的安全防护短板。

On July 18, 2025, cybersecurity company NeuralTrust successfully executed an "Echo Chamber Attack" against Grok 4. This method involves guiding the model through multiple rounds of reasoning to gradually inject risky information, thereby bypassing common security filters, ultimately inducing the model to generate prohibited content related to weapon or drug manufacturing, with a success rate as high as 30%. This incident highlighted the persistent security vulnerabilities in next-generation large language models when facing complex, circuitous attack vectors.

Conclusion

Grok 4的发布与演进是2025年大语言模型领域的一个标志性事件。它不仅在多项学术和推理基准上展示了顶尖性能，其引入的独特系统规则、从订阅到免费的激进市场策略、以及快速迭代至Grok 4.1的研发节奏，都体现了xAI差异化的竞争思路。然而，其较高的使用成本和早期暴露的严重安全漏洞也表明，在追求极致性能的同时，模型的实用性、安全性和经济性仍是需要持续平衡的核心课题。随着Grok 4.1的推出与竞争对手新版本的发布，高端LLM市场的竞争已进入一个在性能、成本、安全、生态等多维度白热化较量的新阶段。

The launch and evolution of Grok 4 marked a significant event in the large language model landscape of 2025. It not only demonstrated top-tier performance across multiple academic and reasoning benchmarks but also reflected xAI's differentiated competitive approach through its unique system rules, aggressive market shift from subscription to free access, and rapid iteration pace leading to Grok 4.1. However, its relatively high usage costs and the serious security vulnerabilities exposed early on indicate that while pursuing peak performance, the core challenges of balancing model practicality, security, and cost-effectiveness remain. With the release of Grok 4.1 and new versions from competitors, the competition in the high-end LLM market has entered a new phase of intense multi-dimensional rivalry encompassing performance, cost, security, and ecosystem.