eWEEK content and product recommendations are editorially independent. We may make money when you click on links to our partners. Learn More.
Elon Musk announced the release of Grok 3, the latest family of chatbots from his xAI company, earlier this week during an X livestream. The billionaire called Grok 3 the “smartest AI on Earth,” but it enters a crowded field of competitors with a headstart. Here’s what you need to know about the new release.
What are the different Grok 3 models?
Musk said the first four models of the Grok 3 family to be launched would be the standard Grok 3, Grok 3 mini, Grok 3 Reasoning, and Grok 3 mini Reasoning. The models were trained on synthetic data and are designed to self-correct by analyzing inconsistencies.
The Grok 3 Reasoning and Grok 3 mini Reasoning models specialize in breaking down their decision-making processes step-by-step and fact-checking themselves before delivering responses, similar to DeepSeek-R1 and o3-mini. These models are best at resolving mathematics, science, and programming queries. The Grok 3 mini versions will generate responses more quickly at the expense of some accuracy checks.
Introducing DeepSearch and other key features
All Grok 3 chatbots are able to respond to questions and generate or analyze images. They also come with a new agentic feature called DeepSearch. Like the Deep Research tools from OpenAI, Google, and Perplexity, DeepSearch will scan the internet to provide a summarized response, citing the sources used. Users can also activate special modes:
- Think mode instructs Grok 3 to engage in a more careful reasoning process and provide a more accurate answer.
- Big Brain mode allocates additional computing resources to tackle very advanced, multistep tasks.
Additionally, Grok 3 has “real-time knowledge of the world via the X platform,” and also powers the social app’s AI assistant, explains posts, generates memes, and more.
How much does Grok 3 cost?
Subscribers to X’s Premium+ tier can access the four Grok 3 models, and the cost has been increased to $40 per month to account for this addition. More advanced features, including DeepSearch, Think prompts, and higher image generation limits, will soon be made available through a new xAI subscription plan called SuperGrok, rumoured to be priced at $30 per month or $300 per year.
How can you access Grok 3?
If you have an X Premium+ account, Grok 3 can be accessed through the Grok website, Grok iOS app, and Grok chat interface on X. A Grok Android app will be released soon, and pre-registration is open through the Google Play Store. Grok 3 and DeepSearch will also be accessible through xAI’s API, which is currently only available in beta.
Not all of the Grok 3 models or features have been released yet and some are still in beta, but Musk confirmed that “voice mode,” which will allow users to speak their prompts and listen to responses, will be available in about a week.
How powerful is Grok 3?
The Colossus Supercluster is an AI supercomputer developed by xAI and located in Memphis, Tennessee. It was recently expanded to contain 200,000 GPUs, and is presumably where Grok 3 was pretrained, utilising at least half of the units.
In an X post, Musk claimed that pretraining it used “10X more compute than Grok 2” leading it to be an “order of magnitude more capable,” as he confirmed in Monday’s livestream. Its dataset has been expanded from Grok 2’s to include legal filings. During the livestream, the xAI team showed a number of charts displaying Grok 3 and Grok 3 mini’s performance relative to Gemini-2 Pro, DeepSeek-V3, Claude 3.5 Sonnet, and GPT-4o on numerous benchmarks. Both of them outperformed all the others in mathematics, science, and coding.
The Grok 3 Reasoning and Grok 3 mini Reasoning were compared to o3-min, o1, DeepSeek R-1, and Gemini-2 Flash Thinking for the mathematics benchmark, and also came out on top. These claims were verified by AI benchmarking platform lmarena, which put it at number one for all its categories.
Concerns over AI knowledge distillation
Musk said that some of the “thoughts” generated by reasoning models will not be displayed to users. This is to prevent distillation, a practice where developers extract knowledge from one AI model to train another — a controversy that DeepSeek was accused of by OpenAI. With more features rolling out soon, it remains to be seen whether Grok 3 truly is the “smartest AI on Earth.”