Elon Musk’s xAI rolls out next-gen AI models Grok 3 and Grok 3-mini with reasoning capabilities

First launched nearly two years ago, Grok AI has been positioned by Musk as an ‘anti-woke’ alternative to foundational AI models developed by other tech giants such as Google. It is known for providing unfiltered responses on political subjects and other sensitive topics, where competitors have implemented stricter safeguards.

The name ‘Grok’ was borrowed from a science fiction novel called Stranger in a Strange Land by Robert A Heinlein. It means “to fully and profoundly understand something,” according to Musk.

Grok 3

Grok 3 was initially slated to be released in 2024. It has reportedly been trained using a massive cluster of 2,00,000 graphics processing units (GPUs) housed in a data centre in Memphis, Tennessee, US. This supercomputer called Colossus was established within a span of six to eight months, according to xAI.

“Grok 3 is an order of magnitude more capable than Grok 2. [It’s a] maximally truth-seeking AI, even if that truth is sometimes at odds with what is politically correct,” Elon Musk said during the launch that was live streamed on X.

https://t.co/HTK4u2aYRu

— Elon Musk (@elonmusk) February 18, 2025

Musk claimed that Grok 3 was developed with “10x” more computing power than Grok 2. xAI has also launched a smaller version of Grok 3 called Grok 3-mini that is capable of responding to questions more quickly.

Grok 3 finished pre-training in January 2025. It was tested across benchmarks that evaluated the AI model, as well as its smaller version, on mathematical skills, code generation, and scientific knowledge.

Grok 3 benchmark testing results. (Screenshot: xAI)

An early version of Grok 3 codenamed ‘Chocoloate’ was also submitted for a blind test in Chatbot Arena, an open platform for crowdsourced AI benchmarking.

‘Blind testing’ results of Grok 3. (Screenshot: xAI)

Grok 3 Reasoning

Similar to reasoning AI models such as OpenAI’s o1, Grok 3 is also capable of ‘thinking through’ problems.

In a live demonstration of Grok 3’s reasoning capabilities, xAI researcher Igor Babuschkin asked the AI model to calculate and plot a trajectory from Earth to Mars and Mars to Earth. It was also prompted on the live stream to develop a new video game that is a mixture of existing games, Tetris and Bejeweled.

While the chain of thought (CoT) of Grok 3 was visible, Musk said that the “thoughts” are somewhat obscured to prevent distillation, a method used by developers to train their AI model on the outputs of other AI models. OpenAI recently accused Chinese AI startup DeepSeek of intellectual property (IP) theft by distilling from its models to create its own.

xAI claimed that a beta version called Grok 3 Reasoning outperformed the best version of o3-mini across several benchmarks, including a new evaluation of high school-level mathematical skills called AIME 2025.

Grok 3 Reasoning and Grok 3 mini Reasoning benchmark testing results. (Screenshot: xAI)

“Because it can reason and think, you can also ask the models to think longer which is called Test-Time compute. In this case, the shaded bar means that we asked the model to spend more time thinking about the problem before concluding what is the right solution,” xAI researcher Jimmy Ba said.

Grok 3 Reasoning and Grok 3 mini Reasoning results for AIME mathematical benchmark. (Screenshot: xAI)

Grok 3 DeepSearch

In addition to the pre-trained and reasoning versions of Grok 3, xAI released a new feature in the Grok app called DeepSearch that is capable of scanning the internet (including X), processing the gathered information, and providing a comprehensive report on the topic with various sections such as initial search and data collection, background and context, etc.

Grok’s new DeepSearch feature comes after rivals Google, OpenAI, and Perplexity released their own AI tools for deep research.

X Premium+ subscribers will reportedly get early access to Grok 3. A new subscription package called SuperGrok on the Grok app and website will also include advanced features such as additional reasoning, DeepSearch queries, and unlimited image generation. It is priced at $30 per month or $300 per year.

However, API access to Grok 3 models and its DeepSearch functionality for enterprises will be rolled out after a few weeks.