Elon Musk’s xAI has officially launched Grok 3, touted as “the smartest AI on Earth”, with a live stream showing its capabilities. The model is on par if not better when compared to leading AI models like OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash Thinking on certain benchmarks.
The model was trained on over 100,000 Nvidia H100 GPUs that provided 200 million GPU hours, which is ten times more than what was used to train Grok 2.
Grok 3 beats GPT-4o on various benchmarks, including AIME (math questions) and GPQA (PhD-level physics, biology, and chemistry questions). It also scored competitively in Chatbot Arena, a crowdsourced test. Grok 3’s reasoning model also outperformed others in the AIME 2024 exam and even a smaller version of the fresh AIME 2025 exam.
The Grok 3 series includes a reasoning model which “thinks” before answering, to provide a more accurate answer. Users can use the “big brain” mode for more difficult questions, which xAi says is best suited for math, science, and coding. The reasoning mode in Grok 3 is also paired with DeepSearch, which as the name suggests, scans the internet and X to analyze information and deliver detailed responses to queries. There’s also Grok 3 mini that could be used to provide quicker outputs.
The reasoning models can be accessed via Grok’s standalone app, although Musk mentioned that some of the model’s “thoughts” are obscured to prevent distillation, where another “child” AI model can distill knowledge from its “parent” model. DeepSeek was recently accused of something similar where it was investigated whether it distilled knowledge from OpenAI’s model to train its own.
Grok 3 will be first available to X’s Premium+ subscribers, which costs $22 a month, while other features are behind a separate upcoming subscription called SuperGrok which would provide access to reasoning capabilities, DeepSearch queries, and unlimited image generation. Musk also said that Grok will soon gain a voice mode as well, and the Grok 3 models will arrive in xAI’s API in a few weeks.
