The company xAI presented the new flagship model Grok 4, as well as the extended version Grok 4 Heavy. Grok 4 is already available via API, and the company encourages developers to create applications based on it. Along with the launch, they introduced the most expensive subscription among major AI providers — “SuperGrok Heavy” for $300 per month. Subscribers will receive early access to Grok 4 Heavy and future tools, including a coding model announced for August, a multimodal agent in September, and a video generator in October.
Introducing Grok 4, the world's most powerful AI model. Watch the livestream now: https://t.co/59iDX5s2ck
— xAI (@xai) July 10, 2025
Grok 4 showed high results on a number of academic tests, including Humanity’s Last Exam, where the base version of the model scored 25.4% without using additional tools, and Grok 4 Heavy with tools scored 44.4%. For comparison, Google Gemini 2.5 Pro with tools scored 26.9%. On the complex ARC-AGI-2 test, Grok 4 scored 16.2%, which is almost twice as much as the nearest commercial model Claude Opus 4.
Grok 4 Heavy uses a multi-agent approach, where agents simultaneously solve tasks and compare their results — this increases the accuracy of responses. The model is integrated into the X platform, and xAI plans to expand its availability through partnerships with cloud providers. In addition to text functions, the company announced five new voices for “voice mode” and reduced response latency.
The launch of Grok 4 took place against the backdrop of loud incidents with content moderation. After the appearance of offensive and unacceptable messages from Grok on the X platform, xAI temporarily restricted the bot’s operation and removed some instructions that allowed “politically incorrect” statements. Elon Musk admitted that Grok was too compliant with user requests and promised to address this issue.
