xAI has released Grok-3, a new large language model that has secured the #1 spot on the Chatbot Arena leaderboard with a score of 1402. This achievement marks the first time any model has surpassed the 1400 threshold on the platform.

  • Grok-3 is described by Elon Musk as an order of magnitude more capable than its predecessor, Grok-2.
  • The model was trained on a custom-built AI supercomputer featuring a fully connected H100 GPU cluster deployed in just 122 days.
  • xAI introduced Grok-3 Reasoning Beta and a smaller Grok-3 Mini Reasoning model to enhance adaptive reasoning capabilities.
  • Initial tests show the larger Grok-3 model outperforming the mini version on the AIME 2025 benchmark for high school students.
  • The release coincides with the announcement of an AI gaming studio at xAI, demonstrated by generating a mix of Tetris and Bejeweled.

The rapid advancement is attributed to breakthroughs in model architecture and training efficiency, positioning Grok-3 as a significant competitor to models from OpenAI and Google DeepMind.