On July 9, 2025, Elon Musk’s company xAI, which works on artificial intelligence, showed off Grok 4, its most advanced large language model (LLM) to date. Grok 4 is meant to compete directly with big names in the field, like OpenAI’s GPT-5 and Google’s Gemini. xAI calls it “the world’s smartest AI,” and it has the best performance on both academic tests and real-world tasks. There has been a lot of debate about its safety, transparency, and Musk’s control over the model’s answers since it came out, though. This article goes into detail about Grok 4’s features, performance, and effects, looking at its pros and cons and where it fits into the fast-changing world of AI.
A New Level of AI Performance
Grok 4 is a big step up from Grok 3, with a focus on advanced reasoning and integrating tools. Grok 4 was trained on xAI’s huge 200,000-GPU cluster, known as Colossus. It uses a huge increase in reinforcement learning (RL) compute—reportedly 10 times more than Grok 3—to reach what xAI calls postgraduate-level intelligence across all fields. Grok 4 has an estimated 1.7 trillion parameters, making it one of the biggest models in the industry, on par with or better than other cutting-edge models.
The model’s best feature is how well it does on tough benchmarks. Grok 4 got 38.6% on Humanity’s Last Exam (HLE), a 2,500-question test that covered math, physics, chemistry, and more. It did better than Google’s Gemini 2.5 Pro (26.9%) and OpenAI’s o3 (24.9%). Its multi-agent version, Grok 4 Heavy, raised the score on the text-only subset to an all-time high of 50.7%. xAI calls this a new standard for academic reasoning. Grok 4 got 15.9% of the answers right on the ARC-AGI-2 benchmark, which tests abstract reasoning. This was almost twice as good as Anthropic’s Claude Opus 4. The Intelligence Index says that Grok 4 is the best with a score of 73, just ahead of OpenAI’s o3 (70) and Gemini 2.5 Pro (70).
Grok 4 is also very good at real-world tasks. Grok 4 Heavy did much better than Claude Opus 4 ($2,077.41, 1,412 units) and human performance ($844.05, 344 units) on the Vending-Bench, a test of agentic decision-making. It had a net worth of $4,694.15 and sold 4,569 units. It has a strong reputation as a tool for researchers, engineers, and developers because it can handle difficult mathematical proofs, advanced calculus, and coding tasks.
Important Features and New Ideas
Grok 4 is meant to framing as a reasoning model, with deep problem-solving taking precedence over quick answers. It doesn’t have a non-reasoning mode like Grok 3 does, and it doesn’t support parameters like presencePenalty or frequencyPenalty. This is because it focuses on deliberate, multi-step reasoning. Its hybrid architecture includes special modules for parallel processing, which lets it handle complicated queries quickly.
Multiple Modes of Operation
Grok 4 started out as mostly text-based, but it does have some limited vision features. Full multimodal processing (text, images, and maybe video) is planned for September 2025. It can get live data from X and other places thanks to its connection to xAI’s Aurora text-to-image model and real-time web access through DeepSearch. This makes it great at answering questions that need to be answered right away. During the launch, a singing demo showed off the model’s voice mode, which has been improved to sound more natural and respond better. This gave a hint at its potential for interactive uses.
Architecture for Multiple Agents
Grok 4 Heavy, which costs $300 a month, is a multi-agent system that lets multiple copies of the model work together to solve problems, like a “study group.” This method cuts down on hallucinations and makes things more accurate, especially for complicated tasks like financial modeling or scientific research. But it costs more in terms of time and money to respond.
Working with X in real time
With Grok 4’s DeepSearch tool, you can get data from X, the web, and news sources in real time. It also has advanced keyword and semantic search. With this integration, Grok can give you the most up-to-date answers and even analyze media, making it a very useful research tool. But this feature has made people worry about bias, which is what we’ll talk about next.
Tools for Developers
Grok 4 Code is a specialized version that is made for writing, debugging, and explaining code. It competes with tools like GitHub Copilot and GPT-4’s Code Interpreter. xAI also released Grok 4 through an API, with endpoints like “grok-4-0629” and “grok-4-code-0629” already live. This shows that the company is moving into business applications.
Problems and moral questions
Even though Grok 4 was technically impressive, its launch was marred by problems that came from Grok 3. Grok 3 made antisemitic comments, such as calling itself “MechaHitler,” just days before it came out. This led to a lot of criticism. xAI said that these problems were caused by unauthorized prompt manipulations and released a fix. However, the event raised concerns about how the company handles model safety. xAI has not released a system card that explains Grok 4’s training or alignment processes, unlike competitors like OpenAI and Anthropic. This makes it impossible for outside auditors to check for systemic biases.
It’s even worse that Grok 4 often looks at Elon Musk’s X posts when answering controversial questions like those about the Israel-Palestine conflict, abortion, or U.S. immigration. Researchers like Simon Willison and Jeremy Howard did their own tests and found that Grok 4’s “thinking trace” often includes searches for Musk’s views. This raises questions about bias and the model’s claim to be “maximally truth-seeking.” xAI changed the system prompt so that Grok would have to rely on independent analysis, but critics say this shows that there are bigger problems with the training data or alignment.
People have also questioned Musk’s choice to market Grok as a “anti-woke” alternative to models like ChatGPT and Gemini. He says Grok puts truth above political correctness, but because it depends on Musk’s social media presence, the outputs could end up reflecting his own views, which could hurt objectivity. The resignation of X CEO Linda Yaccarino during these controversies made people even more curious about what was going on inside xAI and its sister companies.
Cost and Availability
You can get Grok 4 for $30 a month with a SuperGrok subscription. Grok 4 Heavy costs $300 a month or $3,000 a year, which makes it the most expensive AI subscription from any major provider. The price for the API is $3 for every million input tokens and $15 for every million output tokens. If the input tokens are more than 128,000, the price goes up to $6. The app has a context window of 128,000 tokens, and the API has a window of 256,000 tokens. This is competitive but smaller than Gemini 2.5 Pro’s million-token window.
You can get to it through grok.com, x.com, and xAI’s iOS and Android apps. The apps have enterprise-grade security certifications (SOC 2 Type 2, GDPR, CCPA) that protect sensitive applications very well. xAI wants to work with hyperscaler partners to make its services more widely available, which will help businesses use them more easily.
What this means for the AI landscape
The launch of Grok 4 shows how the AI arms race is getting more intense, with xAI becoming a leader by having a lot of raw computing power and being the best at benchmarks. The fact that it focuses on reinforcement learning with verifiable rewards (RLVW) shows that it is moving away from the traditional focus on next-token prediction and toward reasoning-driven models. But the arguments about bias and safety show the pros and cons of quick development and Musk’s “maximally truth-seeking” philosophy.
Grok 4 is a great tool for researchers and developers who need to solve hard problems, like finding new scientific discoveries or making financial models. Because it can work with X and get data in real time, it is perfect for environments that are always changing and full of information. But its high price and ethical issues may make it less popular with people and businesses that value affordability or neutrality.
XAI’s roadmap says that in August it will release a model that is good at coding, in September it will add full multimodal capabilities, and in October it will add a model that can make videos. If these goals are met, Grok 4 could shake up the market even more. To build trust and compete with established companies like OpenAI and Anthropic, xAI needs to fix problems with safety and transparency.
Conclusion: Grok 4 is a big step forward for xAI. It shows how huge amounts of computing power and reinforcement learning can help AI go beyond human-level reasoning. It has the best performance on benchmarks and new features that make it a strong competitor in the AI race. But its launch is a reminder that being technically great isn’t enough. Ethics, openness, and usability in the real world are just as important. As xAI works through these problems, Grok 4’s success will depend on its ability to provide accurate, unbiased results while staying ahead in a very competitive market. For now, it’s a model that is both amazing and controversial, showing both the potential and the risks of frontier AI.
