After months of delay, Elon Musk’s xAI finally launched its newest AI model, Grok 3—packed with new features and capabilities. Basically, Grok did a speed run to catch up or even surpass other AI models like OpenAI’s ChatGPT and Google’s Gemini. Musk even boldly claims it to be the “Smartest AI on Earth.” So, what’s new in Grok 3, and does it live up to the hype? Let’s dive in.
Table of Contents
Grok 3
Now, Grok can better analyze images and tackle more complex problems, especially in math and coding. During a livestream, Musk said Grok 3 is “maximally truth-seeking,” meaning it provides honest answers, even if they are controversial.
Grok 3 isn’t just one model—it’s a family just like other AI models from Gemini and ChatGPT. Alongside the main version, there is Grok 3 Mini, which responds faster but may trade off some accuracy. The main version beats almost all other AI models in benchmarks, surpassing OpenAI’s 4o, Gemini 2.0 Pro, DeepSeek V3, and even Claude 3.5 Sonnet. Meanwhile, the Mini model performs on par with these top-tier models. Both models also performed well in Chatbot Arena, where users vote on AI model responses.
Musk says Grok 3 was built with ten times more computing power than its predecessor, Grok 2. The model was trained using 200,000 GPUs in a massive Memphis data center and learned from an expanded dataset, including court filings.
Grok 3 Reasoning Models
Along with the mainstream Grok 3 models, xAI also introduced reasoning models—Grok 3 Reasoning and Grok 3 Mini Reasoning. In simple terms, a reasoning model is an AI model that solves problems step-by-step, taking its time to reach a more logically accurate answer.
These reasoning models are designed to tackle tricky problems in math, science, and coding by fact-checking their answers more carefully. Users can access the Grok 3 reasoning model by selecting Think mode. Additionally, there is a special Big Brain mode for extra-detailed reasoning on tough questions.
Even the Grok 3 reasoning models surpass all other AI reasoning models in benchmarks. Interestingly, Grok 3 Mini Reasoning performed better than Grok 3 Reasoning. However, the company claims that Grok 3 Reasoning was trained for less time comparatively and has a high potential to improve in future updates. In the demo, the model was able to calculate the spacecraft travel from Earth to Mars and even created games by combining two games.
While Grok lets users check its reasoning process, the company says it performs additional reasoning in the background that it does not show. This is to protect Grok 3’s reasoning process from being copied by competitors, following recent controversies where the Chinese company DeepSeek was accused of copying OpenAI’s models.
DeepSearch
Along with their new AI models and reasoning models, xAI also introduced their first AI agent—DeepSearch. This feature is similar to the Deep Research option on ChatGPT and Gemini. When you enter a research question, instead of providing information from just a few sources, DeepSearch searches dozens or even hundreds of sources to deliver a more detailed report. Rather than performing only a surface-level search, it digs deeper, continuously searching for more information based on what it finds online.
You can use this by just selecting the DeepSearch mode on the Grok. Similar to the reasoning model, you can see the thought process of the AI model with DeepSearch.
Get Started With Grok 3
Grok 3 will be released today, along with the reasoning model and DeepSearch feature. For now, it is available exclusively to X Premium+ subscribers. xAI is also offering a SuperGrok plan for $30 monthly or $300 yearly, which unlocks more reasoning and DeepSearch queries, plus unlimited image generation.
Musk announced that Grok 3 would soon get a voice mode, allowing users to interact through spoken conversations. In the coming weeks, the model will also become available to businesses through xAI’s enterprise API. Additionally, once Grok 3 is stable, xAI plans to open-source Grok 2.
When Musk first introduced Grok two years ago, he promised it would be more direct and less filtered than other AI models. Grok and Grok 2 were known for giving raw, sometimes edgy responses but struggled with political topics, often showing a left-leaning bias. Musk blamed the training data and vowed to shift Grok closer to neutrality. Whether Grok 3 achieves this remains to be seen.
