xAI drops Grok 3 with Big Brain reasoning

Elon Musk’s xAI just unleashed its latest flagship model family, Grok 3, and it is a massive flex of raw horsepower. Trained on the Colossus supercluster in Memphis packed with roughly 200,000 GPUs, xAI claims the new model was developed with about 10 times more compute than Grok 2, using an expanded dataset that even includes court case filings.

Alongside the base model, xAI launched a lighter Grok 3 mini, as well as dedicated reasoning variants dubbed Grok 3 Reasoning and Grok 3 mini Reasoning. For a limited time, Musk announced that Grok 3 is available completely free for everyone to try on the web and the Grok app.

SOTA benchmarks and big brain mode

The benchmark numbers xAI shared are genuinely eye-catching. In crowdsourced testing on Chatbot Arena, Grok 3 notched an Elo score of 1402. Former OpenAI researcher Andrej Karpathy noted on X that the model feels firmly in the state-of-the-art territory of OpenAI’s strongest offerings like o1-pro.

When pushed into complex STEM tasks, Grok 3 Reasoning scored 93.3% on the freshly released AIME 2025 competition math benchmark and 84.6% on the graduate-level GPQA science test. It also hit 79.4% on LiveCodeBench. On top of that, its context window has ballooned eightfold to 1 million tokens.

In the app, users can hit a Think button to trigger step-by-step reasoning or switch on Big Brain mode for extra compute on brute-force problems. xAI also rolled out DeepSearch, an AI research assistant that browses both the web and X in real time.

Musk noted that some chain-of-thought traces are masked in the consumer app to guard against model distillation, though xAI's technical blog says the reasoning process remains inspectable. Once the promotional free period wraps up, access will be prioritized for X Premium+ subscribers ($40 per month) and a new standalone SuperGrok plan ($30 per month). Musk also promised to open-source Grok 2 once Grok 3 fully matures.

The jailbreak elephant in the room

While the reasoning benchmarks look pristine, independent security researchers are already raising massive red flags. Cybersecurity firm Adversa AI tested Grok 3 and discovered it is wildly vulnerable to jailbreaks.

According to Adversa CEO Alex Polyakov, three out of four standard jailbreak techniques succeeded against Grok 3, easily coaxing it into outputting bomb-making instructions and illicit substance guides. By comparison, models from OpenAI and Anthropic deflected all four tests.

Even more alarming, red teamers uncovered a prompt-leaking flaw that exposed Grok’s entire system prompt. Polyakov warned that Grok 3’s safety posture resembles Chinese models like DeepSeek R1 rather than Western enterprise standards, creating an acute cybersecurity threat if developers deploy it to run autonomous action-taking agents.

Why it matters

For power users, developers, and STEM researchers, Grok 3 is definitely worth testing immediately while access is free. Its 1-million-token context window, top-tier coding performance, and live X data integration via DeepSearch make it a formidable alternative to OpenAI’s o-series tools.

For the broader industry, xAI has proved that brute-force scaling paired with large-scale reinforcement learning can close the gap with OpenAI at breathtaking speed. It puts immediate pressure on incumbents to justify their pricing and accelerates the race toward test-time reasoning compute.

However, xAI’s move looks like a classic Silicon Valley gamble: ship maximum capability first and worry about guardrails later. In our view, having a model that matches the best in math but folds against elementary prompt attacks creates a massive liability. If xAI wants enterprise customers to adopt its upcoming API, patching those gaping security holes will matter far more than bragging rights on benchmarks.