xAI has unveiled an early preview of Grok 3, its most advanced model blending strong reasoning with extensive pretraining knowledge, alongside a new AI agent called DeepSearch. Trained on the Colossus supercluster with ten times the compute of previous state-of-the-art models, Grok 3 features significant improvements in mathematics, coding, and instruction-following through large-scale reinforcement learning.
- Grok 3 (Think) utilizes chain-of-thought reasoning for seconds to minutes, achieving 93.3% on the AIME 2025 benchmark and an Elo score of 1402 in Chatbot Arena.
- Grok 3 mini offers cost-efficient reasoning, reaching 95.8% on AIME 2024 and 80.4% on LiveCodeBench.
- The model features a 1 million token context window, eight times larger than previous versions, and state-of-the-art accuracy on the LOFT benchmark for long-context retrieval.
- DeepSearch is introduced as an agent that uses code interpreters and internet access to synthesize information and generate comprehensive reports.
Grok 3 is now available to X Premium and Premium+ users, with API access coming in the coming weeks. The release marks a step toward scalable oversight and adversarial robustness as xAI prepares to train even larger models on its 200,000 GPU cluster.