Following earlier teasers, Elon Musk’s AI venture, xAI, has officially unveiled the next-generation large language model Grok 4, introducing both a single-agent version and a more advanced Grok 4 Heavy variant capable of multi-agent collaboration. The latter supports up to four simultaneous agent tasks and boasts an impressive contextual understanding window of 256,000 tokens, enabling it to tackle extended texts and complex tasks with significantly greater efficiency.
According to xAI, the training scale of Grok 4 is ten times larger than that of its predecessor, Grok 3, and a staggering one hundred times greater than Grok 2. The resources allocated during its reinforcement learning phase reportedly surpass those of industry titans like OpenAI and Google, with xAI claiming Grok 4’s training investment exceeds other models on the market by more than a factor of ten.
Internal benchmark results shared by xAI show Grok 4 outperforming its main competitors in several critical evaluations. In the “Humanity’s Last Exam” benchmark without tool assistance, Grok 4 achieved a 25.4% score, surpassing Google’s Gemini 2.5 Pro at 21.6% and OpenAI’s o3 at 21%. With tool augmentation, Grok 4 Heavy scored an impressive 44.4%, far exceeding Gemini 2.5 Pro’s 26.9%.
In the ARC-AGI-2 benchmark—designed to assess artificial general intelligence—Grok 4 attained an accuracy of 15.9%, nearly double the performance of the second-place Claude Opus 4, underscoring Grok 4’s superiority in reasoning and logical deduction.
In terms of voice interaction, Grok 4 introduces five newly developed natural speech styles, infusing its responses with nuanced emotional expression and more human-like tonality. xAI has also halved the response latency, delivering a more fluid and instantaneous conversational experience.
Aiming to capture the developer and advanced user market, xAI will launch Grok 4 Code in August, promising enhanced code generation and debugging capabilities. In September, Grok 4 will gain multimodal functionality, initially supporting text and image input, with plans to expand into video processing and real-time data retrieval, significantly broadening its application scope.
However, Grok 4 adopts a premium pricing model: API usage costs $3 per 1 million tokens of input and $15 for the same output volume. For access to the multi-agent Grok 4 Heavy, users must subscribe to the “SuperGrok Heavy” plan at $300/month, surpassing the $200 fee for ChatGPT Pro and clearly targeting enterprise clients and professional users. xAI has even showcased its utility in game development workflows to demonstrate its high-end potential.
Meanwhile, the basic version of Grok 3 remains free to use, serving as an accessible entry point for general users.
Related Posts:
- X Blocks AI Training: Musk’s New API Rules & Grok’s Edge
- X (Formerly Twitter) Silently Trains AI on User Data, Sparks Privacy Concerns
- Elon Musk’s xAI Aims for $20 Billion Fundraising
- Microsoft Edge is 29% faster than Chrome?
Support Our Threat Intelligence
If you find our CVE report and cybersecurity news helpful, consider supporting our work.