xAI’ Grok 3 Unveiled: Will It Truly Become “the Smartest”?

On February 17, at 8 pm PT, Elon Musk's startup xAI unveiled Grok 3, their new most advanced AI model, with a live demo of its capabilities. The release positions Grok 3 as "the smartest AI on earth" and a direct competitor to OpenAI's ChatGPT, DeepSeek R-1, Google's Gemini, and Anthropic's Claude. 

Grok 3 is not a single model but a whole series of models that also includes Grok 3 mini, which provides faster answers but with less accuracy. In addition, xAI introduced a separate line of Grok 3 Reasoning and Grok 3 mini Reasoning. These models are able to analyze queries more deeply and independently check the correctness of their answers, similar to "reasoning models" such as OpenAI's o3-mini and DeepSeek's R1. The reasoning models are known for carefully checking themselves before publishing their results, which allows them to avoid some of the pitfalls that traditional models tend to fall into. 

Grok-3 has two reasoning submodes. One model is called "Think," which will display Grok's reasoning as it solves queries. At the same time, Musk noted that some of the “thoughts” of the reasoning models in the program will be hidden to prevent distillation, a technique currently used by AI modelers to borrow data from other models. OpenAI recently suspected that this was how the DeepSeek model was trained, which allegedly took data from ChatGPT. Another submodel is "Big Brain," designed for solving even more complex tasks that involve additional computing power. xAI is also launching a Grok AI agent product called Deep Search similar to the one recently introduced by OpenAI. 

Grok 3 was built with several key improvements over Grok 2. Compared to Grok 2, the new version is significantly faster, resulting in shorter response times and improved overall usability. Unlike previous models, which relied heavily on human-generated data, Grok 3 was trained on synthetic data. This approach allows for greater efficiency and adaptability, as well as a reduction in bias, although at the same time it may raise reliability concerns. To train Grok 3, xAI reportedly used a massive data center in Memphis that houses around 200,000 GPUs. The model was trained on xAI’s Colossus supercomputer, which was built in just eight months. The system uses 100,000 GPU hours from Nvidia, making it one of the most compute-intensive AI projects to date. Musk himself noted that the latest model was built with “10 times” more computing power than Grok 2, and an expanded training dataset that allegedly includes court cases. 

While not officially confirmed, there are speculations that Grok 3 could feature text-to-video capabilities, an area of ​​AI that is rapidly growing. There is also a reason to believe that Grok 3 will support multimodal processing, meaning it will be able to process text, images, and possibly even audio and video data, similar to OpenAI’s GPT-4 Turbo.

According to xAI, Grok 3 outperforms GPT-4o on benchmarks including AIME, which evaluates a model’s performance on a set of math questions, GPQA, which evaluates models on PhD-level physics, biology, and chemistry and LiveCodeBench which evaluates models on their coding abilities. An early version of Grok 3 also reportedly achieved competitive results in Chatbot Arena, a benchmark that pits different AI models against each other and collects user ratings of their answers. Musk also noted that Grok 3 is focused on "seeking the truth," even if it doesn't coincide with the "politically correct" position. This approach aims to counteract perceived biases in other AI models, such as OpenAI's ChatGPT, which Musk has criticized for being "trained to be politically correct.”

Musk's startup also boldly claims that Grok 3 Reasoning outperforms the best version of the o3-mini, the o3-mini-high, on a newer math test AIME 2025.

Grok 3 will initially be available to X Premium+ subscribers ($22/month or $229/year), including in the xAI Grok for Android and xAI Grok for iPhone mobile apps. In addition, xAI is going to launch a new SuperGrok plan for $30/month (or $300/year) that unlocks additional reasoning and DeepSearch capabilities, as well as unlimited image generation.

xAI claims a voice response mode will be coming to the Grok app in the coming week. Additionally, an API will be released for developers to integrate Grok 3 functionality into third-party apps. It can also be expected that Grok 3 will be integrated into many parts of X platform, including the Grok tab, the “explain this post” feature, and more. Musk also confirmed that xAI plans to open source Grok 2 once Grok 3 is stable, which he said could happen within a few months.

Final Thought

Musk's AI ventures with xAI and Grok are ambitious. His claims that it’s “the smartest AI on Earth,” align with his marketing style, while significant computing resources to train the model are clear evidence of the scale of the investments. Will Grok 3 live up to its hype and keep up with OpenAI, Google’s DeepMind, and Anthropic that are also developing their next-gen models? We have yet to see, but it’s clear that such tough competition in the industry is pushing it forward and fueling further innovation.