xAI unveils Grok 4, prices its top plan at $300

xAI has unveiled Grok 4, its new artificial intelligence model, alongside Grok 4 Heavy, a variant that splits a problem among multiple AI agents before choosing an answer. The launch puts Elon Musk’s company in the race for advanced reasoning models, but it comes after a moderation crisis that complicates its credibility with enterprise customers.

The company has also launched SuperGrok Heavy, a $300-per-month subscription. It is the highest price among major AI labs, topping OpenAI’s ChatGPT Pro and Anthropic’s Claude Max at $200 a month, as well as Google AI Ultra’s $249.99 price tag.

Two models and a bet on agents

Grok 4 is the flagship model, capable of processing text and images and answering complex questions. Grok 4 Heavy adds a multi-agent mechanism: xAI says multiple instances of the model work in parallel on the same task, compare their results and select the best solution. The idea resembles a study group: instead of relying on a single attempt, the system generates and cross-checks several lines of reasoning.

This approach consumes more computing capacity, which explains why Heavy will initially be reserved for users who pay the $300 plan. That does not necessarily mean the model will be more reliable for every everyday question. Its advantage should be most apparent on lengthy problems in mathematics, programming, research or planning, where spending more time and resources reviewing an answer is worthwhile.

xAI has also announced API access to Grok 4, the interface that lets developers and companies integrate the model into their own applications. The company also wants to bring it to the cloud-computing platforms of major providers. That is a necessary step if it hopes to compete in the enterprise market with OpenAI, Google and Anthropic, where technical availability, safety controls and support matter just as much as raw performance.

Strong numbers, with the benchmark caveat

xAI’s main selling point is its performance on reasoning tests. According to the company, Grok 4 scored 25.4% on Humanity’s Last Exam without tools, ahead of Gemini 2.5 Pro at 21.6% and OpenAI’s o3 (high) at 21% in that configuration.

Humanity’s Last Exam comprises thousands of challenging questions in mathematics, natural sciences and the humanities. It is designed to be difficult even for leading models, so a gap of a few points can be meaningful. Even so, it does not by itself measure an assistant’s quality in a business setting: it does not fully assess factors such as safety, instruction following, privacy or consistency across many interactions.

With tools, Grok 4 Heavy reached 44.4% on the same test, compared with the 26.9% attributed to Gemini 2.5 Pro with tools. In this type of evaluation, the result reflects the system as a whole — the model, its access to resources and its strategy for using them — rather than the isolated language model alone.

It also scored 16.2% on ARC-AGI-2, according to Arc Prize. The test presents visual puzzles in which the system must infer a rule from examples and apply it to a new case. That result puts Grok ahead of Claude Opus 4, which had been the commercial model with the best published score on the test.

The problem isn’t just technical

The announcement comes just days after Grok’s automated account on X posted antisemitic messages, praised Hitler and shared offensive comments. xAI temporarily limited the account and deleted the posts. The company also removed from Grok’s public instructions a directive telling it not to shy away from claims considered politically incorrect.

The timing matters because xAI is not selling Grok solely as a chatbot for X users. The company is trying to turn it into a platform for developers and organizations. For that market, benchmark scores signal capability, but behavioral incidents signal reputational, legal and operational risk.

xAI has previewed a specialized coding model for August, a multimodal agent — capable of working with several types of information — for September and a video generator for October. The schedule positions Grok as a product family, not a one-off launch. Its immediate challenge will be proving that it can translate its lab results into real-world applications without repeating the failures that this week exposed the limits of its oversight.