xAI has released Grok 3, the latest flagship model in Elon Musk’s AI lineup, and paired the launch with new capabilities for the Grok iOS and web apps. The rollout positions Grok more directly against models like OpenAI’s GPT-4o and Google’s Gemini, while also expanding how users can search, reason, and generate responses through the product.
The launch is not just one model. Grok 3 is a family of models, including a smaller Grok 3 mini version that is designed to answer faster while trading off some accuracy. Some models and features are still in beta or not yet fully available, but xAI began rolling them out on Monday.
A larger training push behind Grok 3
Grok already powers features on X and can analyze images and answer questions. Grok 3 builds on that base after several months of development. It had been optimistically slated for release in 2024, but missed that target.
xAI has been training Grok 3 using an enormous data center in Memphis with around 200,000 GPUs. In a post on X, Musk said Grok 3 was developed with “10x” (or so) more computing power than Grok 2. He also said the model used an expanded training set that includes filings from court cases and more.
During a livestreamed presentation on Monday, Musk described the new model in sweeping terms: “Grok 3 is an order of magnitude more capable than Grok 2.” He also called it a “maximally truth-seeking AI, even if that truth is sometimes at odds with what is politically correct.”
Those claims matter because Grok’s identity has been tied to how it answers difficult or controversial questions. Musk originally pitched Grok roughly two years ago as edgy, unfiltered, and anti-“woke.” Earlier Grok models sometimes lived up to that positioning, including by using vulgar language when asked. But they also hedged on political subjects and would not cross certain boundaries.
Reasoning models become a central feature
The biggest product shift in Grok 3 is the addition of reasoning-focused models. Two models in the Grok 3 family, Grok 3 Reasoning and Grok 3 mini Reasoning, are built to “think through” problems before answering. The approach is similar to reasoning models such as OpenAI’s o3-mini and DeepSeek’s R1.
Reasoning models attempt to check their own work before producing a final answer. That can help them avoid some of the usual problems that trip up AI systems, especially on questions that require several steps.
xAI says these reasoning models are especially suited for mathematics, science, and programming questions. In the Grok app, users can ask Grok 3 to “Think.” For harder queries, they can use “Big Brain” mode, which applies additional computing to the reasoning process.
xAI claims Grok 3 beats GPT-4o on benchmarks including AIME, which evaluates performance on a sampling of math questions, and GPQA, which uses PhD-level physics, biology, and chemistry problems. The company also says an early version of Grok 3 scored competitively in Chatbot Arena, a crowdsourced test where users compare AI model responses.
For the reasoning models specifically, xAI claims Grok 3 Reasoning surpasses the best version of o3-mini, o3-mini-high, on several popular benchmarks, including AIME 2025.
DeepSearch brings research into the Grok app
The new reasoning models also support DeepSearch, a new feature inside the Grok app. DeepSearch scans the internet and X, analyzes information, and returns an abstract in response to a question.
That puts xAI into the same product category as AI-powered research tools such as OpenAI’s deep research. The practical idea is straightforward: instead of only answering from the model’s existing training, Grok can look across current online material and X, then condense what it finds into a response.
Musk also said some of the reasoning models’ “thoughts” are hidden in the Grok app to prevent distillation. Distillation is a method AI developers can use to extract knowledge from other models. The issue has been in the spotlight after DeepSeek was accused of distilling OpenAI’s models to create its own.
Access starts with X Premium+ and SuperGrok
Access to Grok 3 will begin with subscribers to X’s Premium+ tier, priced at $50 per month. Other capabilities will sit behind a new plan xAI calls SuperGrok.
SuperGrok is priced at $30 per month or $300 per year, if leaks are to be believed. The plan unlocks additional reasoning and DeepSearch queries and includes unlimited image generation.
The staggered access model makes Grok 3 both a model launch and a subscription expansion. xAI is using the release to add more advanced functions to the Grok app while also separating heavier-use features into paid tiers.
What xAI says comes next
xAI’s roadmap for Grok 3 includes voice and enterprise access. Musk said the Grok app will gain a “voice mode” in the future, as soon as about a week from now. That feature will give Grok models a synthesized voice.
A few weeks after that, Grok 3 models are expected to become available through xAI’s enterprise API, along with DeepSearch. That would move the model beyond the consumer app and into business-facing use cases.
xAI also plans to open source Grok 2 in the coming months. Musk described the company’s approach this way: “Our general approach is that we will open source the last version [of Grok] when the next version is fully out.” He added that when Grok 3 is “mature and stable,” probably within a few months, xAI will open source Grok 2.
One unresolved question is whether Grok 3 changes the political behavior Musk has criticized in earlier versions. Musk has blamed past behavior on training data from public web pages and pledged to “shift Grok closer to politically neutral.” The launch makes clear that xAI is pushing hard on capability, compute, reasoning, and product features. Whether it has also reached that stated neutrality goal is not yet clear.