Elon Musk's xAI launches Grok 3 with reasoning and DeepSearch

Just as the entire tech world was convinced that cheap open-weight distillation would kill massive compute clusters, Elon Musk answered with 200,000 GPUs.

Late Monday, xAI officially launched Grok 3, its latest flagship AI model, alongside a suite of updates for the Grok web and iOS apps. The launch missed its original 2024 target by a few months, but the hardware backing it makes it impossible to ignore.

According to Musk, xAI trained Grok 3 in its massive Memphis data center using roughly 10 times the compute of Grok 2, feeding it an expanded dataset that even included court filings. In true Musk fashion, he called it a maximally truth-seeking AI during Monday's livestream, claiming it jumps an entire order of magnitude over its predecessor.

Big brains and obscured thoughts

Grok 3 is launching as a model family. Alongside the flagship, there is Grok 3 mini for faster responses, as well as specialized reasoning variants: Grok 3 Reasoning and Grok 3 mini Reasoning.

xAI claims the standard Grok 3 tops GPT-4o on benchmarks like AIME and GPQA. Meanwhile, the reasoning version reportedly edges out OpenAI’s o3-mini-high on newer mathematics benchmarks like AIME 2025.

Inside the Grok app, users can prompt the model to "Think," or trigger a heavier "Big Brain" mode for complex problems that require extra compute.

Notably, Musk mentioned that parts of the model’s internal reasoning traces are intentionally obscured in the interface. The goal is simple: prevent rival labs from distilling Grok’s reasoning steps into cheaper models, a direct nod to recent controversies around Chinese labs mimicking frontier outputs.

DeepSearch and the $50 gate

Grok 3 also powers DeepSearch, xAI's direct answer to OpenAI's deep research. The tool combs both the open web and real-time posts on X to synthesize comprehensive research summaries.

Musk promised a synthesized voice mode within about a week, with enterprise API access following a few weeks later. He also reiterated his open-source rhythm: once Grok 3 is fully stable in a few months, xAI will open source Grok 2.

Getting your hands on Grok 3 right now is expensive. It is rolling out first to X Premium+ subscribers at $50 per month. A standalone plan called SuperGrok is also in the works, reportedly priced at $30 per month or $300 per year for heavier reasoning quotas and unlimited image generation.

Why it matters

For power users, Grok 3 looks like a legitimate alternative to OpenAI's reasoning stack for high-level math, coding, and technical research. But at $30 to $50 a month, everyday users have little reason to jump ship right away.

For the AI landscape, Grok 3 serves as a massive proof point that raw compute scaling is far from dead. By throwing 200,000 GPUs at pre-training and reasoning, xAI has forced its way into the front row alongside OpenAI, Google, and Anthropic in record time.

We think pairing reasoning models with X's firehose of real-time public data via DeepSearch is xAI's smartest move, creating a moat competitors cannot easily replicate. However, hiding reasoning chains to fend off distillation shows how paranoid frontier labs have become about IP leakage. The real question now is whether burning cash on this scale can translate into a sustainable business model, especially with rival labs offering competitive intelligence for a fraction of the price.