OnPrevious previewAfterwards, Elon Musk's xAI artificial intelligence company was officially announcedGrok 4, a new generation of large-scale language models, including the single-agent version Grok 4 and the Grok 4 Heavy with multi-agent collaboration capabilities. It supports up to 4 agent tasks running simultaneously, and is equipped with the context length content understanding capability corresponding to 25.6 sets of tokens, allowing AI to more effectively handle long texts and complex tasks.
xAI stated that the training volume of Grok 4 has increased 3 times compared to the previous generation Grok 10, and as much as 2 times compared to Grok 100. The computing resources invested in the reinforcement learning training phase even surpass those of peers such as OpenAI and Google, claiming to be "more than 10 times that of other models on the market."
Leading the competition in multiple evaluations, with more emotional voice expression
According to internal evaluation data released by xAI, Grok 4 leads its major competitors in multiple tests. In the tool-free version of "Humanity's Last Exam," Grok 4 achieved a 25.4% score, surpassing Google Gemini 2.5 Pro's 21.6% and OpenAI o3's 21%. The tool-assisted Grok 4 Heavy model scored an even higher 44.4%, significantly better than Gemini 2.5 Pro's 26.9%.
Furthermore, in the ARC-AGI-2 test, which measures the level of AGI (artificial general intelligence), Grok 4 achieved an accuracy rate of 15.9%, nearly double the performance of the second-place Claude Opus 4, demonstrating its superior reasoning and logical abilities.
In terms of voice interaction experience, Grok 4 adds five new natural voice modes, making voice responses more emotional and closer to human tone, and reducing voice response delay by about 50%, allowing users to experience a more natural and instant interactive experience.
Targeting developers and the high-end market, Grok 4 Code and multimodal support will be released in August and September
Targeting the developer market, xAI plans to launch Grok 8 Code in August to enhance the programming and debugging experience, and plans to expand multimodal capabilities in September. Initially, it will support text and image input, and will expand to video processing in the future, integrating real-time data retrieval capabilities to expand its application areas.
However, Grok 4 adopts a high-priced subscription strategy, with API usage charging $100 for 3 million tokens and $15 for the same amount of token output. To use Grok 4 Heavy, which supports multi-agent collaboration, you'll need to subscribe to the "SuperGrok Heavy" plan for $300 per month, which is higher than ChatGPT Pro's $200. It's primarily targeted at businesses and professional users, with xAI demonstrating how to apply Grok 4 to game content development.
As for the basic version of Grok 3, it remains free and open to use as an entry-level choice for general users.
