Elon Musk’s AI Company, Xai, released its latest flagship AI model, the Grok 3, late Monday, and announced new features for Grok iOS and web apps.
Xai’s answers to models such as Openai’s GPT-4o and Google’s Gemini Grok can analyze images to respond to questions and provide many features of Musk’s social network, X.Grok 3, for months . It was scheduled for release in 2024, but I missed the deadline.
Monday is an ambitious launch.
Xai uses Memphis’ vast data centers to train the Grok 3 with around 200,000 GPUs. In a post on X, Musk claimed that the Grok 3 was developed with “10 times” (or SO) of computing power than its predecessor, the Grok 2. Includes expanded training sets including filings from lawsuits.
“The Grok 3 has more single digit capabilities than the Grok 2,” Musk said in a live streaming presentation Monday. “(It’s) an AI seeking the most truth, even if it sometimes conflicts with what is politically correct.”
The Grok 3 is, to be precise, a family of models. The Grok 3 Mini, a small version of the Grok 3, responds quickly to questions at the expense of some degree of accuracy. Not all models and related features are available yet (some are in beta), but they began rolling out on Monday.
Xai has released GPT-4O with benchmarks including AIME (evaluating the performance of models regarding mathematical questions sampling) and GPQA (evaluating the models using PHD-level physics, biology and chemistry problems). He claims to hit it three times. According to Xai, the early version of the Grok 3 is a crowdsourcing test in which the Chatbot Arena is competitive, and it pits a variety of AI models against each other and votes for priority responses to users.
Two models in the New Grok 3 family, Grok 3 Reasoning and Grok 3 Mini Reasoning, are a way to carefully “consider” issues, similar to “inference” models such as Openai’s O3-Mini and Chinese AI company Deepseek’s R1. You can do it. Inference models try to fact-check themselves before they produce results. This helps avoid some of the pitfalls that usually make the model trip.
Xai claims that Grok 3 reasoning surpasses the best version of O3-Mini (O3-Mini-High) in several popular benchmarks, including a new mathematical benchmark called AIME 2025.
These inference models are accessible through the GROK app. Users can ask Grok 3 to “think.” Alternatively, for more difficult queries, you can leverage the “big brain” mode for inference using additional computing. Xai describes inference models as being perfect for mathematics, science and programming questions.
Musk said that in Grok apps, the way AI model developers use to extract knowledge from other models, “thinking” in inference models is obscure in Grok apps. Recently, Deepseek was accused of distilling Openai’s models to create his own.
Grok’s inference model supports a new feature in the Grok app called Deepsearch, Xai’s answer to AI-driven research tools like Openai’s Deep Research. DeepSearch scans the Internet and X to analyze information in response to questions and provide abstraction.
X’s Premium+ Tier subscribers ($50 a month) will first get access to the Grok 3, with other features being measured behind the new plan Xai calls Supergrok. SuperGrok unlocks additional inference and DeepSearch queries and throws unlimited image generation at a price of $30 a month or $300 a year (if you believe leaks);
In the future, Musk said the Grok app will acquire “voice mode” about a week from now. A few weeks later, the Grok 3 model will be available through Xai’s Enterprise API and through deep search capabilities.
Xai plans to open Grok 2 in the coming months, Musk said.
“Our general approach is to open source the last version (of GROK) when the next version is completely out,” he continued. “If your Grok 3 is mature and stable, and you’re probably within a few months, I’ll open Source Grok 2.”
When Musk unveiled Grok about two years ago, he pitched the AI model as edgy, filtering and anti-‘awakening’. He told me some of those promises. For example, being told that it is vulgar, Grok and Grok 2 is willing to obligate, spitting out colorful languages that you don’t hear from ChatGpt.
However, the Grok model before Grok 3 is hedged into political subjects and does not cross certain boundaries. In fact, one study found that Grok remains on the left in topics such as transgender rights, diversity programs, and inequality.
Musk has condemned the actions of Grok’s training data (public web pages) and pledged to “get Grok closer to political neutrality.” It is not yet clear whether Xai has achieved that goal and what the outcome will be.
TechCrunch has a newsletter focused on AI! Sign up here to get it every Wednesday in your inbox.