Grok 3, xAI’s latest AI model, represents a significant leap in artificial intelligence, seamlessly integrating advanced reasoning with a vast repository of pretraining knowledge. Developed using the Colossus supercluster, which boasts approximately 200,000 GPUs, Grok 3 was trained with ten times the computational power of its predecessor, Grok 2. This substantial increase in resources has enabled Grok 3 to excel in various domains, including mathematics, coding, world knowledge, and instruction-following tasks.
Grok 3 – artistic impression. Image credit: xAI
Enhanced Reasoning Capabilities
A standout feature of Grok 3 is its sophisticated reasoning ability. Through large-scale reinforcement learning, the model has been trained to engage in extended thought processes, ranging from a few seconds to several minutes. This training allows Grok 3 to evaluate multiple approaches, rectify errors, and provide precise answers. Users can activate this deep reasoning mode by pressing the “Think” button, which offers transparency by revealing the model’s step-by-step thought process.
Benchmark Performance
<p”>Grok 3 has demonstrated exceptional performance across various benchmarks. Notably, it achieved a 93.3% success rate on the 2025 American Invitational Mathematics Examination (AIME) and an 84.6% score on graduate-level expert reasoning assessments (GPQA). Additionally, Grok 3 attained a 79.4% success rate on LiveCodeBench, highlighting its prowess in code generation and problem-solving. For users seeking a balance between performance and efficiency, xAI offers Grok 3 mini, a cost-effective variant optimized for STEM tasks that may not require extensive world knowledge.
| Benchmark | Grok 3 Beta | Grok 3 mini Beta | GPT-4o | Gemini 2.0 Pro | DeepSeek-V3 | Claude 3.5 Sonnet |
|---|---|---|---|---|---|---|
| AIME’24 | 52.2% | 39.7% | 9.3% | — | 39.2% | 16.0% |
| GPQA | 75.4% | 66.2% | 53.6% | 64.7% | 59.1% | 65.0% |
| LCB | 57.0% | 41.5% | 32.3% | 36.0% | 33.1% | 40.2% |
| MMLU-pro | 79.9% | 78.9% | 72.6% | 79.1% | 75.9% | 78.0% |
| LOFT (128k) | 83.3% | 83.1% | 78.0% | 75.6% | — | 69.9% |
| SimpleQA | 43.6% | 21.7% | 38.2% | 44.3% | 24.9% | 28.4% |
| MMMU | 73.2% | 69.4% | 69.1% | 72.7% | — | 70.4% |
| EgoSchema | 74.5% | 74.3% | 72.2% | 71.9% | — | — |
Expansive Pretraining and Contextual Understanding
With a context window of one million tokens—eight times larger than previous models—Grok 3 can process extensive documents and handle intricate prompts while maintaining high accuracy in following instructions. This capability is particularly beneficial for tasks involving long-context retrieval-augmented generation (RAG), as evidenced by its state-of-the-art accuracy on the LOFT (128k) benchmark, which encompasses 12 diverse tasks.
Integration with Grok Agents
To bridge the gap between AI and real-world applications, xAI has introduced Grok Agents. These agents, equipped with code interpreters and internet access, enable Grok 3 to seek additional context, adjust its approach dynamically, and enhance its reasoning based on feedback. The inaugural agent, DeepSearch, is designed to meticulously explore vast information repositories, synthesize key insights, and distill clarity from complex data. This functionality is invaluable for accessing real-time information, conducting comprehensive research, and obtaining detailed summaries that surpass traditional browser searches.
Availability and Future Developments
Grok 3 is currently accessible to X Premium and Premium+ users on platforms such as X and Grok.com. Premium+ subscribers gain immediate access to advanced features, including the “Think” mode and DeepSearch. In the coming weeks, xAI plans to release Grok 3 and Grok 3 mini through their API platform, extending access to both standard and reasoning models. DeepSearch will also be made available to enterprise partners via the API. As training continues, users can anticipate frequent updates, new features, and expanded capabilities, including tool use, code execution, and advanced agent functionalities.
Since the launch of Grok 1 in November 2023, xAI’s dedicated team has propelled significant advancements in AI innovation. With Grok 3, they continue to push the boundaries of core reasoning capabilities, leveraging the expanded Colossus supercluster. For those passionate about contributing to the future of AI, xAI encourages applications to join their dynamic team.
If you are interested in this topic, we suggest you check our articles:
- How Long Will Grok-3 Be Free to Use and What Will the Costly Pricing Be?
- Grok AI – Does Elon’s Chatbot Rival Its Competitors?
- AI Chatbot Assistant for Business – Friend or Foe?
Source: xAI
Written by Alius Noreika
