xAI Unveils Grok 3: Elon Musk's Latest AI Model

Elon Musk's AI outfit, xAI, dropped their latest big deal, Grok 3, late Monday, and they didn't stop there—they also rolled out some fresh features for the Grok iOS and web apps. Grok, xAI's take on AI models like OpenAI's GPT-4o and Google's Gemini, isn't just about text anymore; it can now handle images and answer your questions, plus it's the brains behind a bunch of cool stuff on Musk's social platform, X. Grok 3, which they've been cooking up for a while, was supposed to hit the scene in 2024, but, well, you know how these things go—it's a bit late to the party.

This launch is no small potatoes. xAI's been using a massive data center in Memphis, packed with around 200,000 GPUs, to train Grok 3. Musk took to X to brag that Grok 3 was built with "10x" (give or take) more computing power than its predecessor, Grok 2, and they've thrown in a bigger training set, including court case filings and more.

During a livestreamed event on Monday, Musk was all about how Grok 3 is "an order of magnitude more capable than Grok 2." He's pushing for an AI that's all about the truth, even if it ruffles some feathers. Grok 3 isn't just one model; it's a whole family. There's a smaller version, Grok 3 mini, that's quicker on the draw but might miss the mark a bit more often. Not everything's ready yet—some bits are still in beta—but they started rolling out on Monday.

xAI's claiming Grok 3 is outpacing GPT-4o on some big tests like AIME, which checks how well a model can handle math, and GPQA, which throws PhD-level science questions at it. An early version of Grok 3 also did pretty well in Chatbot Arena, where different AI models go head-to-head and folks vote on who's got the best answers, according to xAI.

In the Grok 3 family, there are two models, Grok 3 Reasoning and Grok 3 mini Reasoning, that take their time to really think through problems, kind of like OpenAI's o3-mini and DeepSeek's R1 from China. These reasoning models double-check their work before spitting out answers, which helps them dodge some common AI traps. xAI says Grok 3 Reasoning is beating the pants off the best version of o3-mini on some key benchmarks, including the new AIME 2025 math test.

You can tap into these reasoning models through the Grok app. Just ask Grok 3 to "Think," or if you're tackling something tough, hit up "Big Brain" mode for some heavy-duty reasoning with extra computing power. xAI says these models are perfect for math, science, and coding questions.

Musk mentioned that some of the reasoning models' inner workings are hidden in the Grok app to stop other folks from copying their homework, a process called distillation. Recently, DeepSeek got called out for allegedly doing just that with OpenAI's models.

The reasoning models are the backbone of a new Grok app feature called DeepSearch, xAI's answer to AI-powered research tools like OpenAI's deep research. DeepSearch scours the internet and X to pull together info and give you a summary in response to your question.

If you're shelling out for X's Premium+ tier at $50 a month, you'll be the first to get your hands on Grok 3. Other goodies will be locked behind a new plan called SuperGrok, which, if the rumors are true, will set you back $30 a month or $300 a year. SuperGrok lets you do more reasoning and DeepSearch queries and gives you unlimited image generation.

Musk also spilled the beans that the Grok app will soon have a "voice mode," letting Grok models chat with you in a synthesized voice. That's coming in about a week, and a few weeks after that, Grok 3 models and DeepSearch will be available through xAI's enterprise API.

In the coming months, xAI plans to open source Grok 2. Musk's approach is to release the last version when the new one's fully out. So, once Grok 3 is all grown up and stable, which should be in a few months, they'll let Grok 2 loose for everyone to tinker with.

When Musk first talked about Grok a couple of years back, he promised an AI that was edgy, unfiltered, and not afraid to stir the pot—willing to tackle controversial questions that other AI systems might shy away from. And he's delivered on that to some extent. Ask Grok or Grok 2 to be a bit crude, and they'll happily go there, unlike the more polite ChatGPT.

But before Grok 3, the models were a bit cautious on political topics and wouldn't step over certain lines. One study even found that Grok leaned left on issues like transgender rights, diversity, and inequality. Musk's pointed the finger at Grok's training data—stuff from the public web—and he's promised to nudge Grok closer to being politically neutral. It's still up in the air whether xAI's pulled that off with Grok 3, and what that might mean down the line.