Facts about the new chatbot Grok 3
Only 17 months after the launch of the first version of the chatbot Grok, the new chatbot Grok 3 is ready. Grok 3 was presented at a live launch webinar on 18 February 2025. Named after a Robert Heinlein novel “Stranger in a Strange land”, Grok means “to fully and profoundly understand something”.
To underline this, Elon Musk explained the purpose of Grok 3: “The mission of xAI and Grok is to understand the nature of the universe, where are the aliens, and whats the meaning of life”. Elon also noted that “A fast AI needs to think like a human. It needs to contemplate all the possible solutions, self-critique, verify all the solutions, backtrack, and also think from the first principle.”
Top 11 Grok 3 ai facts and statistics
- Premium+ price: It will cost $40 pr month or $350 per year in the U.S. Will include basic Grok 3 access, some reasoning and X integrations.
- SuperGrok price: It will cost $30 pr month or $300 per year in the U.S. Will include full power, advanced reasoning, unlimited images and priority updates. Will be standalone platform.
- Free version of Grok: It will be a capped version of Grok 3 with Basic capabilities.
- Release date: 18 february 2025, as announced by xAI and showcased in a livestream with Elon Musk.
- GPUs used in training Grok 3: In phase one, 100,000 H100 GPUs from Nvidia were used. In phase 2, 200,000 H100 GPUs from Nvidia were used.
- Training time phase 1: 122 days.
- Training time phase 2: 92 days.
- Creator of Grok 3: xAI, a company founded by Elon Musk.
- Availability: Accessible now to all X users.
- xAI team size: Approximately 1000 employees.
- Cost estimate: The suggested hardware investment is between $6-8 billion for the 200,000 GPUs. This is excluding the the power, staff, buildings and other expenses.
What is Grok 3 AI
Grok 3 is the latest version of Elon Musk´s chatbot Grok. Its an advanced and highly intelligent chatbot, that offers real-time webbrowsing, super fast performance, advanced reasoning capabilities, and close to human level AI. Its trained on 10 times the computing power of Grok 2, and outperforms many competing LLMs.
Grok 3 can be accessed directly from an X account, or from the website grok.com.
AI Agents in Grok 3 - DeepSearch
The team highlighted DeepSearch as one of the major new products in Grok 3. DeepSearch is a so-called next-generation search engine that allows users to get answers to everyday questions.
As the team stated in the presentation, DeepSearch allows for real-time web browsing and the researching of each webpage, with remembering and cross-validation of different data sources.
The advantage of DeepSearch is that it searches Google, and visits websites real-time. This ensures that its information is up to date.
The team gave a live demo at the launch, and in the demo, DeepSearch can be seen to do real-time searches, analyze and think about the content, cross-validate to other results, and even think about if it needs to ask other questions to reach the result. Just like a human would perform a search.
Deep Search has a feature called Thoughts. It allows users to to see all information that the model thinks is trustworthy and what is not. It also shows how it cross-validates different sources, which makes the search process much more transparent to users.
Think mode in Grok 3 for advanced reasoning and coding
Think mode in Grok 3 is another new feature. When think mode is enabled, it enables the advanced reasoning capabilities of Grok 3. The use cases for this are coding, math problems, and problems that require complex problem-solving capabilities.
Its possible to see how its reasoning to achieve its results. Grok 3 doesn’t reveal all its thoughts, though, because, as Elon Musk remarked: “We are doing some obfuscating of the thinking, so the model won’t get copied instantly.. ”.
The xAI team ran a live demo of the Grok 3 reasoning to demonstrate the Think mode by asking it to solve the physics problem. It thought for 114 seconds and delivered a piece of Python code that ran flawlessly, demonstrating the trajectories of Earth and Mars, and the spaceship, traveling between the two planets. A screenshot of the demo can be seen below.
In an equally impressive live demo, the team created a game designed to be a fusion between Tetris and Bejeweled.
The team notes another fascinating fact about the Think mode. Although Grok’s reasoning capabilities is only trained on math problems and comparative coding problems, it can work on all kinds of other tasks, like creating games. What seems to be happening is that Grok learns this ability to detect its own mistakes and its thinking to correct them, persist in a problem, try lots of different variants, and pick the one that’s best.
So, the generalizing capabilities that Grok learns from math and coding can then be used to solve all kinds of other problems.
What is the performance of Grok 3
During the development of Grok 3, it has been evaluated in 3 categories to test its performance. Math (AIME´24), Science (GPQA) and Coding (LCB Oct-Feb). According to the xAI team, it outperformed all competitors like GPT4o and Germini-2 Pro by a large margin. When the first Grok model launched 17 months ago, it could barely solve any high school problems. And now, its able to graduate from college.
How does Grok 3 perform in Chatbot Arena
In order to gauge real-life performance, the xAI team had Grok 3 running for two weeks on the open-source platform Chatbot Arena. Chatbot Arena is an open and crowdsourced platform designed to evaluate and compare Large Language models.
Chatbot Arena allows human users to interact with two anonymous LLM AI models side-by-side. Users input any prompt they want, get the response from two models, and vote on the best model based on the output.
An early version of Grok 3 scored 1402, beating all other ai models. Compare that to the best model from Openai, ChatGPT-4o-latest, with a current score of 1377, according to the Chatbot Arena LLM Leaderboard from Huggingface.
Training like a boss - the Colossus supercomputer
The supercomputer that was used to train Grok 3 is called Colossus, and it is truly massive. The Colossus has a total number of 200,000 H100 Nvidia GPUS. Elon Musk explained at the launch that they could not build a new building to house the supercomputer due to time constraints, so they had to look for abandoned buildings. They found an Electrolux Factory in Memphis, where the Colossus is placed.
According to Musk, current datacenters told xAI that it would take 18-24 months to make 100,000 GPUs to run coherently. This was the deciding factor behind building the supercomputer “in-house”.
As an anecdote, Elon Musk mentioned that to cool the Colossus, “We leased about a quarter of the mobile cooling capacity in the United States”.
1.2 GigaWatts - Back to the Future Power
At the end of the launch, it was revealed that xAI had begun work on the next computing cluster to replace the current one. The current computing cluster operates on a quarter of a GigaWatt, whereas the next will operate on 1,2 GigaWatt, or as Elon Musk jokingly remarked, “Back to the Future Power.”
When is Grok 3 API coming out
The Grok 3 API is expected to be ready in the coming weeks.
Will Grok 3 be open source
Elon Musk stated that they have a general policy that the previous model will be open-sourced once the current model is stable. He mentioned that they expect that in a few months, once Grok 3 is stable, Grok 2 will be open-sourced. This means we will have to wait for Grok 4 in order for Grok 3 to be open-sourced.
This article written by:
Jesper Nissen has a background in theoretical physics. He got his Masters degree in Physics from the Niels Bohr Institute at the University of Copenhagen. He started working in the IT industry in 2004, and founded his first web agency in 2008. Jesper is the founder of the cloud stacking platform YACSS. He is also the founder of the automated schema platform schemawriter.ai. Jesper is a known public speaker, and have spoken at the most prestigous SEO conference SEO Rockstars. He has also been a speaker at Dark AI, hosted by Holly Starks.