SAN FRANCISCO — xAI has unveiled Grok 3, its most advanced AI model to date, alongside Grok 3 mini—a cost-efficient variant designed to push the limits of reasoning and pretraining knowledge. Developed on the Colossus supercluster with 10 times the compute of earlier state-of-the-art models, Grok 3 combines superior reasoning abilities with extensive pretraining, delivering breakthrough performance in mathematics, coding, world knowledge, and instruction-following tasks.

Grok 3 represents a major leap in test-time compute and reasoning. Using large-scale reinforcement learning, Grok 3 refines its chain-of-thought process, enabling it to think for seconds to minutes. This dynamic reasoning allows the model to correct errors, explore multiple problem-solving approaches, and deliver accurate answers—features that have set new performance standards in academic benchmarks and real-world user preferences. Notably, on the 2025 American Invitational Mathematics Examination (AIME), Grok 3 (Think) achieved an impressive 93.3% score, while Grok 3 mini demonstrated remarkable efficiency on similar STEM tasks.

Equipped with a context window of one million tokens—eight times larger than its predecessors—Grok 3 can process extensive documents and complex prompts without sacrificing accuracy. This vast context capacity, combined with advanced factual accuracy and stylistic control, enables Grok 3 to outperform competitors on diverse tasks. An early version even topped the LMArena Chatbot Arena leaderboard under the codename chocolate, reinforcing xAI’s position at the forefront of AI innovation.

Pioneering a new paradigm in AI interaction, Grok 3 integrates reasoning with practical tool use. With built-in code interpreters and internet access, Grok 3 models can query for missing context, adjust approaches on the fly, and enhance their reasoning through continuous feedback. DeepSearch, xAI’s first AI agent built on this foundation, exemplifies this capability by synthesizing information across vast data sources to provide concise, comprehensive insights—taking AI interaction well beyond a conventional search experience.

Grok 3 and Grok 3 mini are currently available to 𝕏 Premium and Premium+ users on both Grok.com and the 𝕏 app. With advanced features like the Think button, users can inspect not only final answers but the full reasoning process behind them. In addition, xAI is preparing to launch a Grok 3 API that will offer developers low-latency, multi-region access, enhanced security features, and robust management tools. This API rollout is poised to further expand Grok 3’s reach into enterprise applications, setting the stage for a new era of AI-driven innovation.

Since the launch of Grok 1 in November 2023, xAI’s focused team has consistently pushed the boundaries of AI development. With Grok 3, the company is set to redefine what next-generation intelligence can achieve. As training continues and new features—such as multimodal understanding and enhanced agent capabilities—are rolled out, xAI invites both users and developers to join the journey toward building AI for humanity’s future.

In the context of AI, what’s the radical shift here?  

The radical shift lies in integrating advanced, test-time compute with dynamic, multi-step reasoning. Grok 3 leverages reinforcement learning to engage in prolonged, self-correcting thought processes—enabling it to explore multiple problem-solving approaches, handle extensive context, and seamlessly combine reasoning with real-time tool use—marking a new era of transparent, adaptable AI performance.