On February 20, 2025, Elon Musk's xAI released Grok 3—after throwing 10x the compute resources at training compared to Grok 2.

The result? xAI finally caught up to OpenAI and Anthropic on capability.

It only took pouring massive resources into the problem.

What They Trained

Grok 3: The flagship model, competitive with GPT-4o and Claude 3.5 Grok 3 Reasoning: Extended thinking mode competing with o1 Grok 3 Mini: Faster, cheaper option for lighter tasks

Three models covering different use cases.

The 10x Compute Investment

Training Grok 2 used X compute. Grok 3 used 10X.

More H100 GPUs: Expanded supercomputer clusters Longer training: More epochs, more data Better architecture: Improved model design from lessons learned Cost: Estimated hundreds of millions of dollars

Elon's approach: brute force scaling works.

The New Capabilities

"Think" mode: Extended reasoning for complex problems "Big Brain" mode: Maximum compute for hardest tasks DeepSearch: Autonomous web research across multiple sources Tool use: API connections and function calling 128K context window: Long-form content processing

Grok 3 wasn't just faster—it could do things previous versions couldn't.

The X Integration Advantage

Grok 3's killer feature: available to free X/Twitter users.

Distribution: 500+ million potential users Real-time data: Live access to X posts and trends Social features: Share AI conversations, collaborate Free tier: Basic access without subscription Premium benefits: More queries, faster responses for X Premium

No other frontier model had this distribution channel.

The Benchmarks

Grok 3 finally competed on capability:

• MMLU: On par with GPT-4o
• Coding: Competitive programming level
• Math: Strong with reasoning mode
• General knowledge: Finally caught up

Still not beating Claude or o1 on their strengths, but respectably close.

The Delayed Launch

Originally expected January 2025. Shipped six weeks late.

Why the delay?

• Training took longer than expected
• DeepSeek disruption changed competitive landscape
• Safety testing and red-teaming
• Infrastructure scaling challenges

Better late and good than rushed and mediocre.

The Strategic Position

xAI's unique advantages:

• X/Twitter integration: Half a billion users
• Elon's capital: Billions to invest in compute
• Less censored: Fewer content restrictions than competitors
• Open weights: Grok 2 was open-sourced, Grok 3 likely will be
• Real-time data: X platform as training and inference data

Different business model than OpenAI/Anthropic/Google.

Where Are They Now?

Grok 3 proved xAI was serious—no longer Elon's AI side project but a legitimate frontier AI lab. The 10x compute bet paid off in capability, though whether it paid off financially remained unclear.

The X integration meant millions tried Grok 3 who'd never paid for ChatGPT or Claude. Distribution mattered as much as capability.

February 20, 2025 was when xAI officially graduated from "catching up" to "keeping up"—demonstrating that with enough capital, infrastructure, and distribution, new companies could still compete in the AI race, even against entrenched leaders.