Elon Musk introduced Grok 4 last night, calling it the ‘smartest AI in the world’ — what businesses need to know

Days after the controversy surrounding a wave of anti -Semitic responses recently conducted by Chatbot, which AI operates on its social network X (formerly on Twitter), which appears to be unmatched, unaware, and called for “Elon Musk” the latest version of the AI ​​Model family, during the event that was fixed last night last night.

Musk was also published on X: “GROK 4 is the first time, in my experience, that artificial intelligence managed to solve difficult engineering questions in the real world where answers cannot be found anywhere on the Internet or in books. It will improve a lot.”

The new version already includes two distinct models: Groc 4The single thinking model, and Groc 4 heavyA multi -agent system designed to solve complex problems through internal cooperation and synthesis.

Both models are improved for thinking tasks and come with the integration of original tools, empowerment capabilities such as web search, implementation of code, and multimedia analysis.

MUSK and his team in Xai standards indicate that GROK 4 surpasses all current competitors through a set of academic assessments and coding, even when compared to the competitors of an artificial intelligence thinking model, such as Openai O3 and Google Gemini.

However, Xai has not yet released a file The form card Or any official issuance notes for GROK 4 for the public, making it difficult to evaluate their performance and the demands submitted independently during the flow. We’ll update if/when this becomes available.

MUSK and members of his Xai team are participating in the live broadcasting they are dealing with the strict controversy facing Grok during the past week, including many Grok incidents that make anti -Semitic notes or indicate themselves as “mecahitler”, and suggests that people who have Jewish titles should deal with them by Hitler – apparently, apparently, and Genocaust.

The closest to the penis was when he stated: “The thing that I think is the most important for the integrity of artificial intelligence-at least tells me the biological nervous network with the most important thing-is that the search for the maximum truth”, and “we need to make sure that artificial intelligence is a good thing.”

However, Musk did not apologize, nor did he accept the responsibility for the anti -Semitic Grok notes, sexual and conspiracy attack. Here is a copy of the full flow:

During the live broadcast, the team emphasized Grok 4’s ability to think of the first principles, correct its own mistakes and may invent new technologies or reveal new scientific visions.

The presentation also included offers for Grok 4 Heavy, which applies multiple agents to address research level problems through disciplines.

Availability and pricing

Grok 4 is now available through many channels, depending on the user type and the level of subscription:

• API arrival (developers and institutions):
  Grok 4 and Grok 4 Heavy live across Xi API. Pricing is organized as follows:
  • 3 dollars per million input symbols
  • $ 15 per million directing symbols
  • 0.75 dollars per million stored input symbols
  • Prices Double after 128,000 symbols In the window of one context
    API supports text and image inputs, job calls, structured outputs, and a window of 256,000.
• Consumer arrival (via Grok Chatbot and applications):
  Individual users can access Grok 4 through Grok chatbot on xthe Grok app (iOS and Android), and x.comBut only with one of the following subscriptions:
  • Premiumlus: 16 dollars per month
  • Supergrok: $ 300 a month
  • new “Supergrok heavy” The layer, also in the price in $ 300 a monthProvides access to No GROK 4 and Grok 4 heavyMulti -agent alternative.
    (Note: Supergrok and Premiumlus levels may vary in sessions available and use via X and GROK platforms.)
• Timing:
  Grok 4 is available immediately after July 9, 2025, Live broadcast. The limits of temporary arrival were in place during the demonstration, but the full work of subscribers began shortly after.
• Expanding the platform:
  Xai referred to plans to provide Grok 4 through Microsoft Azure Ai FoundryGrok 3 and Grok 3 Mini are currently listed.

For subscription details, users are directed to X.AI/Grok and X Premium support. Below, how to compare the pioneering models of Amnesty International in terms of pricing per million symbols.

Provider and model	Window of context	entrance ($/Mtok)	Inputs stored temporarily	Output ($/Mtok)	Additional notes
Xai – GROK 4/4 heavy	256 K (2 x Price> 128 K)	$ 3.00	0.75 dollars	$ 15.00	Enter image, summary of the job, structural Json (APIDOG)
Openai – O3	200 k	$ 2.00	0.50 dollars	8.00 dollars	50 % deduction of the batch-API (Openai, Openai Help Center)
GPT-4O	128 k	$ 5.00	$ 2.50	20.00 dollars	Vision, sound, tools (Openai)
Anthroproy – Claude Sonit 4	200 k	$ 3.00	0.30 dollars	$ 15.00	50 % discount on output (human) removal
Claude Obus 4	200 k	$ 15.00	$ 1.50	$ 75.00	The pioneer (human)
Google – Gemini 2.5 Pro	200 k (2 x Price> 200 k)	$ 1.25	0.31 dollars	10.00 dollars	Arrival of 75 % cache (Google Ai for developers, Google Cloud)
Gemini 2.5 flash	200 k	0.30 dollars	0.075 dollars	$ 2.50	Google Cloud
Deepseek-deepseek-bureasoner	64 k	$ 0.55 (Miss) / 0.14 dollars (beating)	0.14 dollars	$ 2.19	50-75 % discount outside peak times

Unlike its Grok 3 predecessor, which was released in February, which was separated from the responses of tools on general thinking, GROK 4 was trained with tools from the beginning.

The model integrates capabilities such as implementing software instructions, online search and documents connecting. It also provides Groc 4 heavyA multi -agent system where many interior models work in parallel to create and verify answers.

Grok 4 also includes new Sound It includes expressive outputs with decreased cumin, as well as support for the introduction of text and image, organized outputs and job connection.

The most prominent performance

The independent analysis of the AI ​​Model and Benchaorking Group is mentioned on X that Xai provided him with a copy of GROK 4 (not heavy) at a early time from the general release of registration.

In technical standards, Grok 4 leads the artificial analysis intelligence index with 73, ahead of competitors such as Openai’s O3 (70) and Google Gemini 2.5 Pro (70).

He also recorded the highest grades in:

• GPQA Diamond: 88 %
• Arc-Agi 2: 15.9 %, double the best degree
• The last exam for humanities: 24 % on the text version only, and 44 % with tools
• MMLU-PRO and AIME 2024: 87 % and 94 %, respectively
• Coding and mathematics reviews: Supreme so far on LiveCodebeench, SCICODE, Aime24 and Math-500

Despite its standard success, GROK 4’s Output It stands at 75 symbols per second – a flower of models such as Gemini 2.5 Flash (353) or Openai’s O3 (187), but is still faster than Claude 4 OPUS (66).

The model is characterized 256,000 windows context tokeenWhich is located above the limits of the context of 200 kg of O3 And Claude 4 Sonnet, but less than a million symbols provided by Gemini 2.5 Pro and GPT-4.1.

Cases of use of the real world

XAI made many demonstrations to perform GROK 4 in applied scenarios:

• On the mission of a simulation is called VendingbenchGrok 4 dramatically outperformed other models in long horizon financial planning.
• in Sages InstituteGROK 4 researchers used CRISPR records and detecting new hypotheses.
• in RayExplain the X -ray form with a thicker than peer fuel.
• in Financial sectorA combination of access to data in the actual time making it suitable for prediction and analysis.

It can create the model as well 3D video games With the minimum inputs through independent sources and the incorporation of assets. In addition, capabilities to simulate astronomical physical events have shown using the foundation closer of published research.

Reception and discussion

The industry response was divided into Grok 4, enthusiastically for its performance distributor through the criticism of the delivery of the event and the wider confidence issues.

David Shapiro“Grok 4 now takes his place” is smart enough to actually help in Research Research “… but he only signed with Openai.”

Ethan MackeProfessor in Warton, noticed X: “So GROK 3 had three separate incidents as there seem to be unstable changes in the published system that caused a large-scale moral issue and emergency decline. I do not think that you can make a good launch 4 that cannot be enough that there is enough launch, but there is enough launch, and there may be enough launch, and there may be enough treatment- Really good models there.

Bin HayakThe co -founder and CTO criticized the start of the Ai Ai Raindrop AI (himself as a former musk employee).

Despite the criticism, note the artificial company’s standard analysis: “GROK 4 is now the main AI model”.

Continuous confidence issues

Grok 4 launch comes amid renewable criticism of the previous Grok behavior in consumer deployment operations, especially as a built -in Chatbot in the Musk Social Network, X.

During the July 4 holiday and in subsequent days, Grok created anti -conspiracy responses and its erosion that sparked the audit of system design and governance practices.

I also mentioned my colleague in the project, Michael F. Nunez, Grouk answered questions about the Jewish influence in Hollywood by emphasizing that Jewish executives “control leadership” in major studios and influence the content through “progressive ideologies”, retreat from screaming “, and causes screaming. Time, “a clear signal to the Holocaust.

The conspiratorial and anti -Semitic publication was a prolific production and the Anti -Distort Association (ADL), which is a non -profit anti -U.S. -based anti -U.S. -based, which was published on July 8: “What we see from the hard -line construction is now dodging now. It already rises on X and many other platforms.”

This incident follows a history of the problematic GROK output, including the May 2025 case where Grok robot was integrated into X randomly to the “unrealistic white genocide” in South Africa completely from wrong information. Either way, Xai blamed behaviors for the unveiled employees and stated that it was treated.

Indeed, today, GROK 4 users have noticed the consumer application to make an anti -Zionist and anti -Semitic notes:

As I noticed previously, Musk publicly mentioned on several occasions, he wanted to change Grok to better reflect his personal beliefs and lack of confidence in the main media and approved sources. This makes it a bad source of institutions’ contexts, as these observations can negatively affect users and companies that build a Grok family of models.

My previous recommendation remains: for those in the institution who are trying to ensure the work of their business’s AI products properly and accurately … it is better to avoid Grok unfortunately. Fortunately, there are many other alternatives to choose from.