usatoday24
Xai has launched Grok 4its new artificial intelligence model that is sweeping the most demanding performance tests in the sector. The model exceeds OpenAi, Google and Anthropic proposals in multiple benchmarks. And is that in the absence of knowing OpenAI’s strategy with GPT-5At the moment it has all the ballots of being the most pointer the model. However, Grok continues to drag the usual problems: controversial responses, offensive content and an Elon Musk using the model to subtract credibility.
The numbers speak alone. In the test Humanity’s last examconsidered one of the most difficult to measure abilities of AI, XAI states that Grok 4 has achieved 25.4% Without additional tools, surpassing O3 of Openai (21%) and Gemini 2.5 Pro from Google (21.6%). But it is with Grok 4 Heavy, its multiagente version, where the results shoot: according to the company, reaches 44.4% With ‘tools’, almost doubleing competition.
Image: XAI
In addition, at the benchmark ARC-AGI-2which measures the ability to solve complex visual patterns, Grok 4 has obtained 16.2%practically double the next commercial model. According to Musk“Grok 4 exceeds the doctoral level in all subjects, without exception”, an affirmation that, although it sounds to marketing, is supported by the results obtained.
The revolutionary approach. Grok 4 Heavy works with A “multiple agents” system that work in parallel about the same problem, then comparing their results as if it were a study group. This architecture allows you to climb intelligence according to the available computational power, a concept that could redefine how we understand the AI performance.
Image: XAI
The usual problems. The launch of Grok 4 occurs just after the previous version of the chatbot published anti -Semitic comments In X, even identifying as “Mechahitler” in some answers. XAI had to temporarily withdraw the service and eliminate offensive publications, while countries like Poland They announced complaints before the European Commission and Türkiye blocked access to chatbot. The cause was a modification in the system instructions that allowed the model “not to avoid politically incorrect statements.” Although Xai withdrew that guideline, the damage was already done.
A constant sabotage. Despite these technical advances, Musk continues to condition Grok’s responses in ways that compromise their usefulness. The model performs automatic searches for the opinions of the tycoon in X to respond to controversial issues, turning the alleged search for “truth” into an echo of its creator’s ideas.
This practice, confirmed By experts in AI like Carlos Santana, he once again demonstrates how Musk’s controversial decisions are directly influencing the development of the model. In addition, several researchers already They have managed to avoid Easily model safety barriers, making it generate content on chemical weapons, malicious software, drugs and other sensitive issues through relatively simple jailbreak techniques.
Ethan Mollick, professor at Warton and an expert at AI, Point out The lack of transparency of the company: ‘There is no detailed technical documentation, risk analysis or explanations on how to avoid future incidents’. This opacity makes it difficult for companies to trust Grok for critical applications.
Grok 4 prices. XAI Grok 4 Basic offers at a price of about 30 dollars a month. However, the company has also launched Supergrok Heavy, a subscription of $ 300 per month in which its most advanced model is offered and that directly becomes the most expensive market service on the market.
What is coming. XAI plans to launch a programming model in August, multimodal agents in September and video generation in October. Also will integrate Grok 4 in Tesla vehicles Next week, expanding the scope of AI throughout the Musk ecosystem. The question is whether the company will separate the technical excellence of the media controversies that surround it, or if it will remain hostage of its founder’s impulsive decisions.
Cover image | XAI
GIPHY App Key not set. Please check settings