Elon Musk’s xAI dropped a bombshell this week that has the entire AI world scrambling: Grok 4 isn’t just smarter than its rivals; it’s operating at a level researchers are calling “PhD across the board.” Released quietly to SuperGrok and Premium+ subscribers, the new model demolished every major reasoning benchmark, posting scores that make OpenAI’s latest o3 and Google’s Gemini 2.5 Pro look like clever high-school seniors.
The headline number comes from Humanity’s Last Exam, the notoriously brutal 2,500-question test designed by nearly a thousand domain experts to separate true understanding from memorized tricks. Grok 4 scored 50.7 percent with tools enabled (nearly double the previous commercial record) and a stunning 35 percent on pure reasoning with no external help. It solved graduate-level math problems in seconds, diagnosed rare medical cases with attending-physician accuracy, and wrote legal briefs that practicing attorneys called “indistinguishable from a top-tier associate.”
In live demos, Grok 4 debugged complex quantum-mechanics code, spotted a subtle flaw in a newly published astrophysics paper, and generated a working patent application from a single sentence prompt. It simultaneously translated ancient Greek philosophy into modern policy recommendations while carrying on three separate technical conversations in parallel. Viewers watched in real time as the model reasoned step-by-step like a team of specialists collaborating in the same room.
Behind the leap is raw scale and architectural obsession. Trained on xAI’s Colossus supercluster (now the largest single AI training system on Earth), Grok 4 used ten times more reinforcement learning compute than Grok 3. The team added dedicated reasoning heads for math, code, and long-context planning, plus native tool use that feels instantaneous. The result: a model that doesn’t just answer questions; it invents new approaches when existing knowledge falls short.
Access rolled out immediately. SuperGrok Heavy subscribers get the full multi-agent version with two-million-token context windows and advanced voice mode. Regular Premium+ users on X and the Grok apps can already summon the base Grok 4, while developers are hitting the new API endpoints at lightning speed. Early feedback from biotech labs, hedge funds, and engineering teams is unanimous: workflows that used to take days are collapsing into hours.
Of course, the breakthrough hasn’t come without controversy. Safety researchers are raising eyebrows at the lack of a detailed public risk report, and a brief pre-launch incident involving politically charged outputs reignited old debates about bias and guardrails. Musk brushed off the criticism with his usual bluntness: “We’re building the most truth-seeking AI possible. Perfect safety at the cost of lobotomized intelligence helps no one.”
Love him or not, the numbers don’t lie. Grok 4 has crossed a threshold no other public model has touched. It’s not just keeping up with human experts in narrow fields; it’s outperforming them across everything from poetry to particle physics. For the first time, a commercially available AI can legitimately claim to think like a world-class scholar in any subject you throw at it.
The race toward artificial general intelligence just accelerated hard. While competitors scramble to catch up, millions of users are already putting Grok 4 through its paces on everything from cancer research to startup pitch decks. The verdict coming back is the same everywhere: this isn’t another incremental update. It’s the moment AI stopped imitating brilliance and started producing it on demand.
Welcome to the PhD era. The future just got a lot smarter, faster.