On Monday evening, Elon Musk’s AI firm, xAI, launched its new flagship AI model, Grok 3, and introduced enhanced functionalities for the Grok iOS and web applications.
Grok, xAI’s counterpart to OpenAI’s GPT-4o and Google’s Gemini, is equipped to analyze images and answer questions, also powering various features on Musk’s social platform, X. Grok 3, which has been under development for several months, was initially scheduled for a 2024 launch, but ultimately missed that target.
The launch on Monday is a bold step forward.
To train Grok 3, xAI has utilized a massive data center in Memphis, housing approximately 200,000 GPUs. In a post on X, Musk claimed Grok 3 was built with around “10x” more computing capability than its predecessor, Grok 2, utilizing an expanded dataset that includes court case filings, among other data sources.
“Grok 3 is significantly more capable than Grok 2,” Musk stated during a livestream announcement on Monday. “[It’s a] maximally truth-seeking AI, even if that truth occasionally contradicts political correctness.”
Grok 3 consists of a range of models, including a quicker version named Grok 3 mini, which answers queries faster, albeit with a potential accuracy trade-off. Although not all models and features of Grok 3 are currently available (some are still in beta), they began to roll out on Monday.
xAI asserts that Grok 3 outperforms GPT-4o on several benchmarks such as AIME (for assessing model performance on math questions) and GPQA (for evaluating PhD-level physics, biology, and chemistry problems). An early iteration of Grok 3 also generated competitive scores in Chatbot Arena, a crowdsourced evaluation that pits various AI models against each other based on user preferences, according to xAI.
Two models within the Grok 3 family, Grok 3 Reasoning and Grok 3 mini Reasoning, can methodically “think through” challenges, akin to reasoning models like OpenAI’s o3-mini and DeepSeek’s R1. These reasoning models aim to fact-check their responses before delivering results, which helps them sidestep some common misunderstandings typically encountered by models.
xAI contends that Grok 3 Reasoning outperforms the leading version of o3-mini — o3-mini-high — across various notable benchmarks, including the newly established mathematics benchmark AIME 2025.
These reasoning models are accessible through the Grok app. Users can request Grok 3 to “Think,” or — for more complex inquiries — utilize “Big Brain” mode for reasoning that utilizes additional processing power. xAI mentions that the reasoning models are designed for tackling mathematics, science, and programming questions.
Musk indicated that certain “thoughts” from the reasoning models are concealed in the Grok app to avoid distillation, a technique that AI model developers use to extract knowledge from other models. Recently, DeepSeek faced allegations of distilling OpenAI’s models to forge its own.
The reasoning models support a new feature in the Grok app called DeepSearch, xAI’s equivalent to AI-driven research tools like OpenAI’s deep research. DeepSearch examines the web and X to synthesize information and produce an abstract in response to user inquiries.
Subscribers to X’s Premium+ tier ($50 monthly) will be the first to access Grok 3, with additional features available through a new plan called SuperGrok. This plan is expected to cost $30 per month or $300 per year (according to rumored leaks), and it offers further reasoning and DeepSearch queries, along with unlimited image generation.
In the near future — potentially within a week — the Grok app will feature a “voice mode,” Musk mentioned, which will enable the Grok models to utilize synthesized speech. A few weeks later, Grok 3 models will be made available through xAI’s enterprise API, along with the DeepSearch function.
Musk stated that xAI intends to open source Grok 2 in the coming months.
“Our general strategy is to open source the last version [of Grok] once the next version is fully released,” he elaborated. “When Grok 3 reaches maturity and stability, likely within a few months, we will open source Grok 2.”
When Musk introduced Grok nearly two years ago, he marketed the AI as edgy, unfiltered, and anti-“woke,” generally willing to engage with contentious topics that other AI platforms avoid. He has followed through on parts of that vision; for instance, when prompted to use vulgar language, Grok and Grok 2 complied, featuring expletives you probably wouldn’t hear from ChatGPT.
However, Grok models preceding Grok 3 avoided addressing political topics directly and refrained from crossing certain lines. Notably, a study indicated that Grok tended to reflect left-leaning views regarding issues like transgender rights, diversity initiatives, and social inequality.
Musk has attributed this behavior to Grok’s training data, which consists primarily of publicly available web content, and has committed to steering Grok toward a more politically neutral stance. Whether xAI has accomplished this objective remains uncertain, along with the potential ramifications.
TechCrunch offers an AI-centered newsletter! Subscribe here to receive it every Wednesday.