‘The Ultimate Truthful AI’: Elon Musk’s xAI Releases Flagship AI Model Grok 3

Elon Musk-founded xAI has unveiled its flagship Grok 3 AI model, along with updates to the Grok iOS app and web version. Grok 3 has been in development for months, with its launch originally scheduled for 2024 having been delayed. Grok 3 was trained with 10 times the computing power of its predecessor, allowing the new AI model to significantly improve its accuracy and insight into data.

Image source: xAI

Grok 3 is the third generation of the xAI family of AI models, which was created to counter developments such as OpenAI’s GPT-4o and Google’s Gemini. The new AI model is a major technological step forward: improved algorithms, increased volumes of training data, the ability to analyze images, and even integration of some features into the X social network. “Grok 3 is an order of magnitude more powerful than Grok 2. It is the most truthful AI possible, even if that truth sometimes diverges from the politically correct,” Musk said during the presentation.

To train Grok 3, xAI used one of the world’s largest data centers, located in Memphis. It uses about 200,000 graphics processing units (GPUs), which allowed it to process more complex data sets and perform calculations at unprecedented speeds. According to Musk, the resources used to train Grok 3 were 10 times greater than those required for Grok 2. In addition, the training set included not only publicly available data, but also court case materials, which potentially expands the capabilities of the new AI model in the field of legal document analysis.

The xAI data center where Grok 3 was trained has 200,000 GPUs, and the expansion from 100,000 to 200,000 GPUs took 92 days.

It is important to emphasize that the new version of Grok is not a single AI model, but a whole family of neural networks adapted to various use cases. For example, Grok-3 mini Reasoning is capable of processing requests at high speed, but this reduces accuracy. However, not all versions of Grok 3 became available immediately – some functions remain in beta testing, but their deployment will begin today.

XAI claims that Grok 3 has shown excellent results in tests, particularly outperforming GPT-4o. It has shown outstanding results in the AIME benchmark, which measures mathematical ability, and the GPQA, which measures PhD-level knowledge of physics, biology, and chemistry. Moreover, an early version of Grok 3 has achieved high rankings in Chatbot Arena (LMSYS), a platform where users compare answers from different AI models and vote on the best ones.

In the Chatbot Arena ranking, an early version of Grok 3, codenamed Chocolate, has shown the highest result among many large language AI models

One of the key innovations was the introduction of Grok-3 Reasoning and Grok-3 mini Reasoning, specialized AI models that can deeply analyze problems, similar to “reasoning” models such as OpenAI’s o3-mini and China’s DeepSeek’s R1. These neural networks don’t just provide answers, but also carefully check facts before formulating them, which significantly reduces the likelihood of errors. According to xAI, Grok-3 Reasoning outperformed o3-mini-high in a number of popular benchmarks, including AIME 2025 Performance.

Grok 3 performance in AIME 2025 tests shows that the Grok-3 Reasoning Beta version outperforms competitors including the o3-mini-high and Deepseek-R1

Users can interact with Grok 3 through the Grok app, which has two modes: Think, for standard queries, and Big Brain, for complex calculations and logical problems. Big Brain uses more processing power to achieve higher accuracy in answers. It is optimal for scientific research, mathematical modeling, and programming. According to Musk, Grok hides some of the AI’s “thoughts” during the reasoning process to prevent distillation, a method used by developers of competing AI models to extract knowledge from other neural networks.

Grok 3 and its mini version outperformed competitors in math, science, and programming tests, beating GPT-4o, Gemini-2 Pro, and DeepSeek-V3

Another important innovation was the appearance of DeepSearch, a tool built on the basis of “thinking” AI models. It performs intelligent searches across open sources on the Internet and data from the X social network, analyzing arrays of information and forming compressed analytical summaries. This functionality makes DeepSearch an analogue of OpenAI Deep Research, but with a more integrated approach to data processing. At the moment, access to Grok 3 is provided to subscribers of X Premium+, the subscription cost is $ 22 per month. In addition, xAI launched a new SuperGrok tariff, which costs $ 30 per month or $ 300 per year. It includes advanced capabilities for reasoning queries, deeper analysis through DeepSearch and unlimited image generation.

DeepSearch in action in the Grok 3 interface, where the system analyzes and searches for relevant information about the upcoming launch of SpaceX’s Starship

Grok will receive an update in the coming week that will add a voice mode, allowing Grok to communicate with users using a synthesized voice. Grok 3 will then be available via the xAI enterprise API in a few weeks, allowing companies to integrate DeepSearch into their business processes. Musk said his company plans to open source Grok 2: “Our approach is that we open source the latest version [of Grok] when the next version is ready. When Grok 3 is mature and stable, which will probably be in the next few months, then we will open source Grok 2.” This means that once Grok 3 is fully stable, developers will be able to study the source code of its predecessor.

Grok was initially positioned as an advanced and alternative AI, capable of freely discussing topics that other neural networks avoid. Research has shown that before Grok 3, the AI ​​model exhibited a political bias, particularly on issues of diversity and inequality. Musk attributed this to the fact that the training data included publicly available web pages that reflected certain ideological positions. Musk promised that Grok 3 would be more politically neutral, but it is unclear whether xAI has achieved this goal.