Every few months another AI lab announces their model topped some benchmark and the discourse goes in circles. Grok-4 is different and it's worth explaining why.
Grok-4 launched in July 2025 and it currently leads every major benchmark for mathematical reasoning and coding. Not by a small margin on one test — across the board, consistently, including evaluations that OpenAI and Anthropic's internal teams designed. That's not a marketing claim. That's what independent researchers are finding when they run the numbers.
The architecture underneath Grok-4 is trained on Colossus — 220,000 NVIDIA GPUs in Memphis running continuously. For context, that's more dedicated AI training compute than most countries have access to in total. When you have more compute than everyone else and you hire the engineers who built the previous generation of frontier models at other labs, you tend to get better models.
Grok Heavy is the tier that most people aren't talking about enough. At $300 per month it's positioned as a professional and research tool — not a consumer chatbot. The people paying $300 a month for Grok Heavy are researchers, engineers, and organizations running complex scientific and technical workloads. That's a different customer than the ChatGPT Plus subscriber. The revenue per user is dramatically higher and the churn is dramatically lower.
The GPT-5 comparison is interesting because OpenAI has been quiet about its release timeline. What's known is that Grok-4 currently leads on the evaluations that matter most for technical work — and that Grok-5 is already in training on Colossus 2 in Mississippi. OpenAI is racing a competitor that has more compute, a faster training cycle, and a distribution platform with hundreds of millions of daily active users built in through SpaceXAI.
The benchmark wars matter this time because the gap between models is real and the infrastructure advantage SpaceXAI has built is not something you close in a product cycle. You close it by building a Colossus of your own. Nobody else is doing that.
Grok-4 isn't winning because of a clever algorithm tweak. It's winning because SpaceXAI built the biggest training cluster on Earth and staffed it with the best people. That's a durable advantage.