Musk's xAI drops Grok 4 with a $300 Heavy tier

Just hours after X's chief executive Linda Yaccarino resigned, and only days after Grok 3 faced intense backlash for spitting out antisemitic responses, Elon Musk took the stage anyway. In typical fashion, xAI answered internal chaos with brute force, officially rolling out Grok 4.

The new release marks xAI's latest reasoning-optimized flagship, accompanied by an advanced sibling dubbed Grok 4 Heavy. It is the first major launch since X was absorbed into xAI in a $33 billion all-stock deal earlier this year.

The Colossus cluster delivers

Grok 4 owes much of its sheer muscle to Colossus, xAI’s massive supercomputer cluster in Memphis. After coming online in September with 100,000 GPUs, xAI expanded the facility beyond 200,000 chips in May, with an eventual goal of one million.

That hardware push shows up clearly in benchmark tables. Testing on Humanity’s Last Exam, a benchmark spanning 2,500 complex scientific questions, Grok 4 solved over 44% of the problems without touching any external search tools. OpenAI's Deep Research tool, powered by o3, posted a 26.6% mark.

Musk claimed on X that the model essentially never misses math or physics exam questions unless they are skillfully adversarial, adding that it can spot ambiguous phrasing and patch errors inside prompts. He even floated that it might start discovering new physics before the year is out.

A $300 monthly tier for multiple agents

The standard Grok 4 is available through the $30 monthly SuperGrok plan, featuring voice, vision, and a 128k context window. Developers can tap the API with a 256k window and integrated search at $3 per million input tokens and $15 per million output tokens.

For power users, xAI introduced Grok 4 Heavy, which uses multiple AI agents to process queries simultaneously, cross-check reasoning, and return the best overall output. That capability lives behind a steep $300 per month SuperGrok Heavy tier, which includes early access to upcoming tools.

Musk also outlined a rapid roadmap: a code-optimized variant next month, an upgraded multimodal version in September, and a dedicated video generation model shortly after that. Eventually, xAI wants to plug Grok into Tesla-style computational fluid dynamics and distribute it across public clouds.

Why it matters

For researchers and technical builders, Grok 4 is a legitimate counterweight to OpenAI’s reasoning reign. The math and physics benchmark leaps are undeniable, making the API compelling right now for anyone wrestling with complex technical workflows or heavy logic pipelines.

However, the $300 monthly tag for Grok 4 Heavy clearly signals an enterprise or well-funded developer target, not a casual consumer product. Most everyday users will find little justification for that price.

Strategically, xAI is showing that raw infrastructure speed can compress competitive gaps fast. By scaling Colossus at breakneck speed, Musk is putting immediate pressure on Google and OpenAI to match his compute scaling.

We think the multi-agent consensus approach in Grok 4 Heavy is a smart way to squeeze higher accuracy out of complex prompts. But the company’s blind spot remains trust. Launching a hyper-capable reasoning model right after an embarrassing safety meltdown means xAI will be scrutinized twice as hard on guardrails and real-world reliability.