Elon Musk drops Grok 4

Just hours after X CEO Linda Yaccarino resigned and following intense backlash over Grok 3's unhinged updates, Elon Musk took to a Wednesday night livestream to drop his biggest AI release yet: Grok 4 and Grok 4 Heavy.

Musk claimed the new reasoning-focused models are better than PhD levels in every subject, even suggesting the AI could discover new physics before the year is out.

Brute-forcing benchmarks with multi-agent design

xAI rolled out two distinct models. The standard Grok 4 is a single-agent model packed with vision, voice, and a 128K context window, available for $30 a month under the SuperGrok tier.

Then there is Grok 4 Heavy, which carries a staggering $300 monthly subscription. Instead of relying on a single chain of thought, Heavy runs multiple agents on the same prompt, has them cross-check their work, and surfaces the best answer.

Developers can also plug into Grok 4 via API with a 256K context window and native web search, priced at $3 per million input tokens and $15 per million output tokens.

On the benchmark front, xAI is claiming substantial victories. On Humanity's Last Exam, a benchmark spanning 2,500 expert-level science questions, Grok 4 scored over 44% without external web search. For perspective, OpenAI's o3-powered Deep Research scored 26.6%. It also posted top scores on Arc-AGI-2 and AIME, edging out Gemini 2.5 Pro and o3.

The horsepower behind this is Colossus, xAI's Memphis supercomputer cluster that went live with 100,000 GPUs last September, doubled to 200,000 in May, and is heading toward 1 million.

Musk also laid out an aggressive roadmap: a coding-focused Grok next month, an expanded multimodal version in September, followed shortly by video generation, and eventually integrations with Tesla's vehicle engineering simulations.

Why it matters

For researchers and heavy technical builders, Grok 4 presents a compelling alternative for deep reasoning and code debugging. A 44% pure-compute score on expert scientific tests suggests it can handle genuine problem-solving rather than surface-level regurgitation. But for most users, the $30 and especially the $300 price tags make this an expensive experiment unless your workflow directly translates reasoning hours into revenue.

For the AI landscape, xAI has proven that rapid hardware scaling can still catch up to incumbent giants. Leaning on the brute force of Colossus, Musk has turned xAI into a legitimate tier-one contender alongside OpenAI and Google in record time, putting immense pressure on rivals to speed up their next-gen reasoning models.

Yet we think xAI's multi-agent approach highlights a growing trade-off. Running multiple internal agents to cross-verify answers is essentially burning compute to artificially inflate benchmark reliability, as reflected in that steep $300 price tag.

More critically, governance remains the elephant in the room. Rolling out Grok 4 right after Grok 3's racist and antisemitic outbursts, paired with the sudden departure of Linda Yaccarino, means xAI is flying at top speed with minimal organizational guardrails. Enterprise buyers will almost certainly hesitate until xAI proves its models can behave consistently under pressure.