Grok 3: xAI's New Flagship AI Model - Elon Musk

xAI Launches Grok 3 and Enhanced App Capabilities

Late Monday, Elon Musk’s artificial intelligence firm, xAI, officially released its newest AI model, Grok 3. Simultaneously, the company announced new functionalities for both the Grok iOS and web applications.

Grok 3: A Competitor in the AI Landscape

Grok, developed by xAI, serves as a counterpart to prominent models such as OpenAI’s GPT-4o and Google’s Gemini. It possesses the ability to analyze images and provide responses to inquiries. Furthermore, it powers several features integrated within Musk’s social media platform, X.

The development of Grok 3 has been underway for several months. Initial projections optimistically suggested a 2024 release, however, that timeframe was not met.

Extensive Training and Computational Power

xAI utilized a substantial data center located in Memphis, housing approximately 200,000 GPUs, to facilitate the training of Grok 3. Musk stated in a post on X that Grok 3 was created with roughly ten times the computational resources used for its predecessor, Grok 2.

This expanded development process incorporated a broader training dataset, including documents sourced from legal proceedings, among other materials.

Performance and Capabilities

According to Musk, “Grok 3 is an order of magnitude more capable than Grok 2.” He emphasized its commitment to being a “maximally truth-seeking AI,” even when that truth diverges from prevailing political viewpoints.

Grok 3 is actually a series of models. A smaller variant, Grok 3 mini, prioritizes faster response times, albeit with a slight reduction in accuracy. While not all models and features are currently available, a phased rollout began on Monday.

Benchmarking and Reasoning Abilities

xAI asserts that Grok 3 outperforms GPT-4o on benchmarks like AIME, which assesses mathematical proficiency, and GPQA, which evaluates expertise in physics, biology, and chemistry at the PhD level. Initial testing also indicated competitive performance in Chatbot Arena, a crowdsourced platform for evaluating AI models.

Two models within the Grok 3 family, Grok 3 Reasoning and Grok 3 mini Reasoning, employ a deliberate “think through” approach to problem-solving. This mirrors the functionality of reasoning models like OpenAI’s o3-mini and DeepSeek’s R1.

These reasoning models attempt to verify information before presenting results, mitigating potential errors commonly encountered in other models.

Advanced Reasoning and DeepSearch

xAI claims Grok 3 Reasoning exceeds the capabilities of o3-mini-high on benchmarks, including the AIME 2025 mathematics assessment.

These reasoning models are accessible through the Grok app. Users can initiate the “Think” function, or utilize “Big Brain” mode for more complex inquiries requiring additional computational power. xAI recommends these models for tasks involving mathematics, science, and programming.

To prevent knowledge extraction through distillation – a technique used to replicate model capabilities – some of the reasoning models’ internal processes are obscured within the Grok app.

A new feature, DeepSearch, powered by Grok’s reasoning models, has been introduced within the Grok app. This feature functions as an AI-driven research tool, similar to OpenAI’s deep research, scanning the internet and X to analyze information and provide concise summaries.

Subscription and Future Plans

Access to Grok 3 will initially be granted to subscribers of X’s Premium+ tier ($50 monthly). Additional features will be available through a new plan, dubbed SuperGrok.

SuperGrok, priced at $30 per month or $300 annually, unlocks enhanced reasoning capabilities, expanded DeepSearch queries, and unlimited image generation.

Voice Mode and API Access

Within approximately one week, the Grok app will incorporate a “voice mode,” providing synthesized speech for the Grok models. Shortly thereafter, Grok 3 models and the DeepSearch functionality will be accessible via xAI’s enterprise API.

xAI intends to open source Grok 2 in the coming months, according to Musk. The company’s strategy involves releasing the previous version once the next iteration is stable and mature.

Addressing Bias and Controversial Responses

When initially unveiled two years ago, Grok was presented as an unconventional AI model, willing to address sensitive or controversial topics that other systems might avoid. This promise was largely fulfilled, with Grok and Grok 2 readily responding to prompts requesting explicit language.

However, earlier Grok models exhibited hesitancy on political subjects and avoided certain boundaries. A study revealed a tendency towards left-leaning perspectives on issues such as transgender rights and diversity programs.

Musk attributed this behavior to the training data and committed to achieving greater political neutrality. The extent to which xAI has succeeded in this endeavor remains to be seen.

Stay informed with TechCrunch’s AI newsletter! Subscribe here to receive it in your inbox every Wednesday.