xAI Launches Grok 4

Elon Musk's AI company xAI officially launched its next-generation flagship model, Grok 4, on Wednesday evening, accompanied by a new premium subscription tier called SuperGrok Heavy, which is priced at $300 per month—making it the highest-priced consumer AI plan among leading vendors.

In recent years, Grok has positioned itself as a strategic alternative to models like OpenAI's ChatGPT and Google's Gemini, with increasingly tight integration into the social media platform X (formerly Twitter). Following xAI's acquisition of X, this alignment has drawn significant attention to Grok's performance in both content generation and real-time public discourse.

In the evening’s livestream, Musk, sporting a leather jacket, appeared with the xAI leadership team to walk viewers through Grok 4’s new features. He described the model as having surpassed PhD-level proficiency in a wide range of academic fields. As for its occasional lack of everyday reasoning or novel insights, Musk downplayed the concerns, stating, “That’s only a matter of time.”

With this release, xAI unveiled two distinct versions of the Grok model. The standard Grok 4 supports functions such as image analysis and Q&A interactions, serving as the company’s flagship offering. Meanwhile, the more powerful Grok 4 Heavy operates on a “multi-agent” architecture, where several AI agents simultaneously address a problem, evaluate each other’s responses, and collectively determine the most accurate result—a mechanism inspired by collaborative learning models.

According to xAI, this structure is a major factor behind Grok 4 Heavy’s high accuracy in handling complex scenarios.

Despite the technological leap marked by Grok 4’s debut, xAI has found itself amid significant controversy.

On the same day as the product launch, X Lab CEO Linda Yaccarino announced her resignation, concluding a nearly two-year tenure.

The company was already under scrutiny just days earlier when Grok’s automated account sparked backlash by inadvertently responding to an anti-Semitic post with a message that appeared to support Hitler. The incident triggered widespread criticism, prompting xAI to quickly limit the account’s access and remove the offensive content.

In the aftermath, xAI discreetly rolled back a recent system prompt that had encouraged the AI not to shy away from “politically incorrect” remarks—a move that many saw as a contributing factor.

Still, during the official Grok 4 presentation, Musk and the leadership team chose not to acknowledge the controversy, keeping the focus squarely on the model’s technical strengths and performance milestones.

xAI reported that Grok 4 delivered impressive results across several industry-recognized benchmarks, including the rigorous “Humanity’s Last Exam” — a comprehensive evaluation designed to test AI systems on thousands of real-world questions spanning mathematics, science, and the humanities.

• Grok 4 (no tools) achieved a 25.4% accuracy rate, outperforming both Gemini 2.5 Pro at 21.6% and OpenAI’s o3 at 21%.
• Grok 4 Heavy (with tools) significantly raised the bar with a 44.4% score, also ahead of Gemini 2.5 Pro’s tool-assisted score of 26.9%.

In a separate advanced benchmark, ARC-AGI-2, which evaluates an AI’s ability to solve visual reasoning and pattern-recognition problems, Grok 4 also set a new record. It earned a score of 16.2%, nearly double that of Claude Opus 4, the next-best commercial model.

To accompany the launch of Grok 4 Heavy, xAI also launched a new high-end subscription service, SuperGrok Heavy, priced at $300 per month, which is currently one of the most expensive individual subscription plans among mainstream AI service providers. Compared with the premium services of platforms such as OpenAI, Anthropic and Google, SuperGrok Heavy is more radical in both pricing and positioning, and can even be said to be an "experimental attempt".

This subscription plan is aimed at professional users who have extremely high demands for cutting-edge performance. Subscribers will enjoy a series of priority benefits, including early access to Grok 4 Heavy models and the first experience of new features that are about to be launched. This means that some features will be tested and iterated on a small scale in this subscription tier before the official release, and users will also have the opportunity to participate in feedback and interaction at the first time.

xAI also announced the update rhythm of future products, drawing a clear product roadmap for subscribers:

1. In August, an AI coding model designed for developers and engineers will be launched to enhance code generation and collaboration capabilities;

2. In September, an intelligent system that supports multimodal input such as text, images, and voice will be launched to bring an interactive experience closer to human perception;

3. In October, the video generation model will be officially released, completing an important part of xAI in the field of content generation.

In addition to personal subscription services, xAI is also actively deploying the enterprise market. Grok 4 has been opened to developers through APIs and supports the construction of multiple application scenarios. Although the enterprise department was established only two months ago, it has planned to cooperate with large institutions through cloud platforms to accelerate the commercialization of Grok.

Although Grok 4 has made breakthroughs in multiple AI benchmarks and has won the attention of many developers with its model capabilities, its controversial brand image problem has not been fundamentally resolved.

For enterprise users, technical strength is important, but the stability, security, and consistency of values of the system cannot be ignored. If xAI really wants to become a strong competitor to ChatGPT, Claude, and Gemini, it needs to be more stable in its development pace and show a stronger response when questioned by the outside world.