February 18, 2012 - Today at 12:00 p.m. Beijing time.Muskits artificial intelligence company xA Musk said Grok 3 is "an order of magnitude" more capable than its predecessor, Grok 2, and is an AI that "strives for the ultimate in realism," even if that realism is sometimes at odds with "political correctness". "political correctness".
In terms of performance of competencies such as mathematical reasoning, scientific logical reasoning and code writing, theGrok-3 Grok 3 has outperformed DeepSeek-v3, GPT-4o, and Gemini-2 pro in many benchmark tests. Musk has even praised Grok 3 as "the smartest AI on the planet".
xAI claims that Grok 3 outperforms GPT-4o in several benchmarks, including AIME (which evaluates a model's performance on math problems) and GPQA (which tests a model's performance on PhD-level physics, biology, and chemistry problems). Additionally, an earlier version of Grok 3 excelled in Chatbot Arena, a crowdsourced testing platform that pits different AI models against each other, with users voting on the better answer.
1AI notes that Grok 3 is not a single model, but a family of models. One of the smaller versions, Grok 3 mini, is capable of answering questions faster at the expense of some accuracy. Not all model versions are currently available online.
The Grok 3 development cycle was reportedly shortened significantly thanks to its powerful Colossus supercomputer. The Colossus supercomputer, which reportedly took only eight months to build, provided strong support for the development of Grok 3. Grok 3 used 100,000 NVIDIA H100 GPUs (later scaled up to 200,000), with a cumulative total of 200 million GPU hours, which is ten times the size of its predecessor, Grok 2. This massive deployment of computing power allows Grok 3 to process massive datasets in less time while significantly improving model accuracy.
The xAI team has not only upgraded the hardware, but also optimized the software. grok 3 further improves the performance of the model by improving the training process, introducing synthetic datasets, self-correction, and reinforcement learning. The combination of these techniques makes Grok 3 even better at handling complex tasks.
two variant versions of Grok 3 -- Grok 3 Reasoning and Grok 3 mini Reasoning.Ability to "think" about problems as carefully as "reasoning" models like OpenAI's o3-mini and DeepSeek's R1. Inferential models perform thorough fact-checking before giving results, thus avoiding some of the errors that usually plague models.
xAI also claims that Grok 3 Reasoning outperforms o3-mini high, the best version of o3-mini, in several popular benchmarks, including a new math benchmark called AIME 2025. xAI also claims that Grok 3 Reasoning outperforms the best version of o3-mini in several popular benchmarks, including a new math benchmark called AIME 2025. Users can access the reasoning model through the Grok app and use the "Big Brain" modes for deeper, more careful reasoning when encountering more difficult problems. xAI says these modes are best suited for math, science, and programming-related problems.
However, xAI also noted thatGrok 3's inference model is still in the testing phase (Beta), and is still being trained!In addition, Grok 3 introduces a new feature called "DeepSearch". In addition, Grok 3 introduces a new feature called "DeepSearch".The company describes it as a new type of search engineDeepSearch is capable of scanning the Internet and the X-Platform for information and responding to user queries in a summarized form.
Musk previously released a video explaining the mission of xAI and Grok -- to understand the nature of the universe. However, the voice mode that was planned for this release did not go live as expected. Musk confirmed this on the X platform, explaining, "There are some issues with voice mode right now and it's expected to launch in about a week, but it's brilliant."
Premium + subscribers to the X platform will be the first to experience Grok 3, while other features are integrated into a subscription service called SuperGrok from xAI. SuperGrok subscriptions are priced at $30 per month or $300 per year, and give users access to additional inference and DeepSearch querying privileges, as well as unlimited image generation capabilities SuperGrok is available in a subscription service called SuperGrok.
Musk also revealed thatGrok will be launching "Voice Mode" in the coming week.and integrating Grok 3 models and DeepSearch capabilities into xAI's enterprise APIs in a few weeks.
also,xAI Plans to Open Source Grok 2 in a Few Months. Musk said, "Our overall strategy is to open source the previous version after the next version is fully launched. When Grok 3 is mature and stable, probably in the next few months, we will open source Grok 2."
The Grok 3 launch comes at a time of escalating rivalry between Musk and OpenAI. Not only has the conflict included lawsuits and wars of words, but more recently there has been an unsolicited $97.4 billion takeover bid for OpenAI by Musk.
