Elon Musk’s startup xAI presented the «smartest artificial intelligence on Earth» — the chatbot Grok 3. More than 100,000 people watched the demonstration of the launch of a new player in the AI market.
The Grok 3 launch went live. xAI claims that Grok 3 outperforms GPT-4o in tests such as AIME, which evaluates a model’s performance on a sample of math questions, and GPQA, which tests models on PhD-level physics, biology, and chemistry questions. According to xAI, an early version of Grok 3 also scored highly in Chatbot Arena, a crowdsourcing test in which different AI models compete against each other and users vote for the best answers.
During a Q&A session to celebrate the Grok3 launch, xAI answered some questions from the audience about the latest AI model. One of them was whether Grok would be a girl or a boy, to which Musk replied, «Grok is whatever you want him to be,» before adding, «People are going to fall in love with Grok. That’s 1,000% likely.»
Who is Grok 3 available to?
«Today, we’re launching Grok3, starting with Premium Plus subscribers on X, who will be the first group to get access. We’re also launching a separate subscription called Super Grok for loyal fans who want the latest features and early access to new features. It’s available on both the Grok app and our new website, grok.com,» Musk said at the presentation.
For $30 per month or $300 per year, SuperGrok also unlocks additional considerations and queries in the new DeepSearch search engine, and adds unlimited image generation.
Two variations of Grok 3, Grok 3 Reasoning and Grok 3 mini Reasoning, can carefully «think through» problems, similar to «reasoning» models such as OpenAI’s o3-mini and R1 from the Chinese company DeepSeek.
Reasoning models carefully test themselves before producing results, which helps them avoid some of the pitfalls that models typically face.
xAI claims that Grok 3 Reasoning outperforms the best version of o3-mini — o3-mini high — on several popular tests, including the new AIME 2025 math test.
How Grok 3 was created
The term Grok comes from Robert Heinlein’s novel Stranger in a Strange Land. It is used by a character who grew up on Mars, and it means a complete and deep understanding of something. The word «grok» conveys a deep understanding, and empathy is an important part of that understanding.
The idea of creating the smartest AI came to Musk in April of last year. In his opinion, the only way to succeed and create the best AI is to create your own data center. In 12 days, the team built a data center and launched the first 100,000 GPUs. Soon, in 92 days, the capacity was doubled. This means that Grok 3 xAI uses a huge data center in Memphis, which contains about 200,000 GPUs, for training.
The chatbot will run on xAI’s Colossus supercomputer, which reportedly uses over 100,000 hours of Nvidia GPUs to train AI models. The system was built in just over eight months.
Moreover, xAI plans to enter into a $5 billion deal with Dell for AI servers for Grok 3, equipped with NVIDIA GB200 semiconductors. This acquisition is intended to meet the growing computational demands of Grok 3 and other AI applications.
What challenges does the smartest AI face?
Grok 3 is a family of models, or rather, not a single model. A scaled-down version of the Grok 3, the Grok 3 mini, answers questions faster, but at the cost of some loss of accuracy. Not all models are available yet, but they begin rolling out on Monday. «This is the most truth-seeking AI ever, even if that truth sometimes runs counter to what is politically correct,» Musk said.
The developers want him to build a viable trajectory for a flight from Earth to Mars and then back to Earth. This involves some complex physical laws that Grok will have to understand. He needs to come up with a viable trajectory, calculate it, and then build it, visualize it.
Grok 3’s preliminary training was completed about a month ago, and since then the team has been hard at work integrating reasoning capabilities into the current Grok 3 model. However, it’s still in the early stages, and the model is still in the training phase. What’s presented today is just a portion of the Grok 3 reasoning model. In parallel, the developers are also training a mini version of the argumentation model.
Reasoning models can be accessed through the Grok app.
Users can ask Grok 3 to «think» or — for more complex questions — use the «Big Brain» mode for additional, more in-depth reasoning. xAI describes these modes as best suited for questions related to math, science, and coding.
Development plans
In a few weeks, Grok 3 models will appear in the xAI enterprise API, along with the DeepSearch feature. And a few months after that, xAI will release Grok 2 as open source, Musk said.
«Our general approach is that we will release the latest version of [Grok] open source when the next version is fully ready,» Musk said. «When Grok 3 is mature and stable, which will likely be within a few months, we will open source Grok 2.»
He promises users a «more polished version» in a week. «We’re also working on a voice interaction feature so you can have conversations. I tried it today and it works pretty well, although it still needs some work. Our goal is to make it so you can talk to it just like you would to a human. I think it will be one of the best experiences using Grok3, but that will happen in about a week,» Elon Musk said.
One of the most humanistic aspects of Grok3 is its creativity. The AI has developed a concept for a new game that brilliantly combines elements of Tetris and Puyo Puyo, illustrating its potential in areas that go beyond science and technology — in entertainment and art. According to Musk, Grok3 is creating a mix of Tetris and Bejeweled: «We are launching an AI game studio called xAI. If you are interested in joining us and developing AI games, please come join us. We will announce the launch tonight.»
«A model that is better than Grok3 must surpass all aspects of the science and engineering of deep learning, so this is no easy task. This will be the last cluster we build and the last model we train. However, we have already started work on the next cluster, which will be about five times more powerful. Instead of a quarter of a gigawatt, it will be about 1.2 gigawatts,» the xAI team says of the next Grok3 cluster.
9ou6ps