Elon Musk’s startup, XAI, yesterday presented its new family of artificial intelligence models, called Grok 3. He presumed that these models surpassed their competitors, but it does not seem that we have a specially striking qualitative leap here. And that is a problem, because Musk wants us to pay for using Grok 3.
To start, Grok 3. In that family the first prominent are Grok 3 Beta and Grok 3 Mini, the two initial versions that directly compete with generative AI models such as GPT-4O, Claude 3.5 Sonnet, Deepseek V3 or Gemini 2.0 Pro. According to the benchmarks used in XAI, Grok 3 surpasses them all, and that good performance seems to be endorsed by independent entities as LMarena.
There is “reasoning” mode. Curiously, Grok 3 does not have a separate version to “reason”, but the model itself is able to activate that mode of reasoning to compete with Deepseek R1 or with O1. Once again XAI reveals that benchmarks in this area seem to place Grok 3 as the one that best behaves.
But reason in silence. Grok 3’s reasoning mode, yes, does not show what you are thinking, something that other reasoning models do. According to XAI, that design decision allows you to prevent competence models from using these data to distill and reuse them in their own developments. It is a zancadilla to the competition, without more.
Deepsearch, the other great protagonist. One of the most striking functions of Grok 3 is Deepsearch, which is nothing more than the response to research tools such as Deep Research, from Openai. Experts like Andrej Karpathy They have tried it And it seems to behave well, but it is actually on horseback among the models of reasoning and de -presence, which is more advanced.
Xai is updated, and that has merit. If a few weeks ago we warned that Grok 3 did not arrive and Xai was staying behindthings have just changed. If we listen to those first demos and internal tests, Grok 3 is a model of the good or better than its current competitors. That is an achievement for Xai, which It was created In April 2023 and has managed to recover all the land that its competitors had already traveled.
It will have a voice. At the moment Grok 3 does not offer the ability to talk through the voice as they do competitors such as Openai or Google models, but Musk indicated that they are working on it. “The voice mode is still a bit irregular,” commented in x“so it probably launches within a week, but it is impressive.”
Grok 3 is not free. Another surprise. After the arrival of Deepseekcompetition does not stop offering advanced options and even reasoning models free of charge. To use Grok 3 of the former you will have to pay the premium+subscription, which costs $ 50 per month. In XAI they have created a new plan called Supergrok that according to rumors will cost $ 30 per month and will also use this model later. Grok 3 will not be initially available in Europe.
And neither do they open source. Although Musk already published the initial version of Grok as Open Sourcehe does not seem willing to do the same immediately with Grok 3. Nor did he do it with Grok 2, although just during the launch of Grok 3 announced that he would soon license that second version as Open Source. As with the decision to hide “how” Grok 3 thinks in its mode of reasoning, Musk seems clear that he prefers not to share too much about his AI models to prevent competition from taking advantage of his progress.
A complicated business model. Precisely the fact of offering it only as part of a payment subscription complicates things to make Grok 3 popularize. A priori it does not seem to be a model of AI that represents a qualitative leap important enough to justify that someone pays for him when there are free alternatives, so, why should we do it?
Musk cult. XAI’s strategy and Elon Musk attracts attention because he seems to force us to choose Grok 3 more for our ideology – Musk’s fans or not – than for a practical and rational issue – which his model is really better and more useful. Benchmarks seem to point to this model behaving something better than competition, but the jump, we repeat, does not seem notable and some independent tests show that Grok 3 is far from perfect, for example When programming.
It already happens with Tesla. It is something similar to what is happening with Tesla and its actions: in recent years the value of the company has shot in the stock market. Its market assessment (1.14 billion dollars) is exceptionally high. As much as to be equivalent to that of the following 16 companies that manufacture cars. Among them are Toyota, Byd, Ferrari, GM, Porsche, Mercedes-Benz, Volkswagen, Ford, BMW, Honda, Hyundai or Stellantis. There is no rational or logical economic reason here: only the popular belief that Tesla “has to be worth more.” It is the cult of Musk made investment (or bet).
Grok 3 will have much to demonstrate. Grok 3 may end up being available for the standard accounts of X. It is what ended up occurring with Grok 2, a model that demonstrated, for example, being especially striking by Not having censorship. However, with Grok 3, the arguments to pay do not seem to be clear, and the demos that were shown in the presentation were basically “one of the same.” And with that it will be difficult for Xai and Musk to convince us that using their AI is the best option.
Image | DVIDS | XAI
In Xataka | Elon Musk lights the “most powerful the world’s cluster.” One with 100,000 gpus nvidia h100