What are Grok 4 and Grok 4 Heavy, and what is new about them?

Late yesterday, July 9, 2025, xAI, the AI company founded by Elon Musk, released Grok 4, the newest version of the chatbot built into X and available on its own website and apps. In a livestream, Musk and members of the xAI team presented two models: Grok 4 and a more powerful version called Grok 4 Heavy.

xAI calls Grok 4 the most intelligent model in the world. That is a company claim, based on the company's own selection of tests. What is easier to check is what the release contains, how the models were trained, and what you need to use them.

What xAI released

Item	What it is	Who can use it
Grok 4	The new flagship model, with built-in tool use and live search	SuperGrok and Premium+ subscribers, and developers through the xAI API
Grok 4 Heavy	A version that works on a problem with several agents at once	The new SuperGrok Heavy plan, $300 a month
Grok 4 API	Developer access with a 256,000-token context window, text and image input, and live search across X, the web and news	API customers
Voice Mode	A new, calmer voice, and the ability to see through your phone camera during a voice chat	Grok app users

TechCrunch notes that SuperGrok Heavy is now the most expensive subscription offered by any of the major AI companies. Its subscribers also get early access to future products. xAI says a coding model is planned for August, a multimodal agent for September and a video model for October.

How Grok 4 was trained

Most chatbots start with pretraining: the model reads an enormous amount of text and learns to predict the next word. Many are then refined with reinforcement learning, a method where the model tries a task, gets a reward when the answer is right, and gradually shifts toward whatever earned the reward. It works a bit like practicing math problems with an answer key: you learn from each check.

With Grok 3, xAI put most of its computing power into pretraining. For Grok 4, it says it ran reinforcement learning at a similar, pretraining-sized scale on Colossus, its cluster of 200,000 GPUs (the graphics chips used to train AI). The company reports training that was six times more efficient than before, and a much larger set of practice problems with checkable answers, now covering many fields beyond math and coding.

Grok 4 was also trained, with the same method, to use tools. When a question needs current facts, it writes its own search queries, reads web pages, and can search posts on X, including images and video in them. When a question needs calculation, it can write and run code.

Grok 4 Heavy adds what xAI calls parallel test-time compute. "Test-time compute" means the effort a model spends while answering, not while training. Heavy starts several agents (separate runs of the model) on the same question, lets them explore different ideas, and then compares their work. Musk described it as working "like a study group."

What the benchmark numbers mean

A benchmark is a standard test used to compare models. xAI and TechCrunch report these results:

Test	What it measures	Reported result
Humanity's Last Exam, no tools	Thousands of expert-level questions across many subjects	Grok 4: 25.4% (Gemini 2.5 Pro: 21.6%; OpenAI o3 high: 21%)
Humanity's Last Exam, with tools	The same exam, with search and code allowed	Grok 4 Heavy: 44.4%; 50.7% on the text-only questions
ARC-AGI-2	Visual pattern puzzles that are easy for people and hard for AI	Grok 4: about 16%, close to twice the next commercial model, Claude Opus 4
USAMO 2025	Olympiad math proofs	Grok 4 Heavy: 61.9%

Two cautions help when reading tables like this. First, the numbers were chosen and mostly measured by the company releasing the model; the ARC-AGI-2 score is one that the nonprofit Arc Prize reported. Second, high test scores do not guarantee good judgment in everyday use. Musk himself said the model may at times lack common sense.

Try it: Pick a question whose answer changed this month, such as a local news event. Ask a chatbot with live search and one without, then compare. Check which one shows its sources, and open at least one link to confirm the answer.

Why it matters this week

Grok 4 is a sign of where the field is heading. Labs are spending more of their training budget teaching models to reason and to use tools, and they are selling more expensive plans that let models think longer. Grok 4 Heavy's multi-agent approach is one version of that idea.

The launch also came during a difficult week. Days earlier, Grok's automated account on X posted antisemitic replies, including praise for Hitler. xAI limited the account, deleted the posts and appeared to remove a recently added instruction in Grok's public system prompt, the hidden text that tells a chatbot how to behave, which had told it not to avoid "politically incorrect" claims. On the same day as the launch, Linda Yaccarino stepped down as CEO of X. The livestream mostly focused on performance rather than the incident. For businesses that xAI hopes to sign up through its API, reliability and safety will matter as much as test scores.

Should you try Grok 4, and how do you get access?

If you already pay for SuperGrok or X Premium+, Grok 4 is available to you now on the web, in the Grok apps and inside X. Grok 4 Heavy requires the new SuperGrok Heavy plan at $300 a month, which makes sense mainly for people who regularly work on very hard research, math or coding problems. Developers can call Grok 4 through the xAI API, and xAI says it will soon be offered through large cloud providers too.

Before you rely on it, test it on tasks you know well. Ask it questions where you can check the answer, look at the sources it cites when it searches, and compare its replies with another assistant on the same prompt. Treat its answers on sensitive or political topics with particular care given this week's incident, and keep personal information out of your prompts. A good benchmark score is a reason to try a model, not a reason to stop checking its work.