What it is
Grok is a family of AI models made by xAI, the artificial intelligence company Elon Musk announced on 12 July 2023. You can talk to Grok on its own website and apps, inside the social network X (formerly Twitter), and in some Tesla cars, and developers can build it into their own software through an API, a way for one program to use another. Like GPT, Claude, and Gemini, it is a language model: a system trained on enormous amounts of text that writes its replies one small piece at a time.
xAI said at its launch that its goal was to "understand the true nature of the universe". It released the first Grok on 3 November 2023 as an early preview for some paying X subscribers. The company's announcement described Grok as designed to answer questions "with a bit of wit" and with "a rebellious streak", and said it was modelled after The Hitchhiker's Guide to the Galaxy, the comic science fiction novel. The name itself comes from science fiction: "grok" is a word coined by Robert Heinlein meaning to understand something deeply. That promised personality, less cautious and more irreverent than rival chatbots, has been Grok's selling point and the source of many of its problems.
The announcement also claimed a feature that still sets Grok apart: "real-time knowledge of the world via the 𝕏 platform". Grok can search posts on X as they are published, as well as the wider web. That makes it quick on breaking news. It also means it reads a stream full of rumours, jokes, and deliberate falsehoods, a point this page returns to.
The generations, at a high level, using xAI's own announcements where possible:
- November 2023: Grok-1, in preview on X for Premium+ subscribers.
- March 2024: xAI released the Grok-1 weights publicly (see below).
- Spring 2024: Grok-1.5, with a much longer context window, the amount of text it can consider at once.
- August 2024: Grok-2 and Grok-2 mini (announced 13 August), which also brought image generation to Grok through an outside model called Flux.
- December 2024: Aurora, xAI's own image generator inside Grok, and a free version of Grok for all X users, with limits.
- February 2025: Grok 3, which xAI said was trained with ten times the computing power of earlier leading models, with a "Think" mode for step-by-step reasoning (see Extended Thinking) and a research tool called DeepSearch.
- July 2025: Grok 4 and Grok 4 Heavy (9 July), which xAI said were trained to use tools such as search and code on their own.
- November 2025 to September 2026: a steady series of point releases, including Grok 4.1 (November 2025), Grok 4.20 (early 2026), Grok 4.3 (spring 2026), Grok 4.5 (July 2026), Grok 4.6 (August 2026), and Grok 4.7 on 21 September 2026, which xAI calls its most capable model for coding and knowledge work.
There are also specialised models, such as a fast coding model and voice models. This line-up will change, probably within months, so check xAI's release notes for the current list.
Open weights for older models. "Open weights" means the company publishes the trained numbers that make up the model, its parameters, so anyone with the right hardware can download and run it. On 17 March 2024, xAI released the weights of Grok-1: a 314-billion-parameter mixture of experts model, the raw pre-trained version rather than the chat version, under the Apache 2.0 licence. That is a true open-source licence that allows commercial use with very few conditions. In August 2025, xAI released the weights of the model Musk called Grok 2.5, but under a custom community licence, not an open-source one. At the time of writing that licence allows research and commercial use as long as you follow xAI's acceptable use policy, requires a "Powered by xAI" notice, and forbids using the model or its outputs to train other large AI models. The files are about 500 GB and need eight powerful graphics cards to run, so this is not something for a home computer. Musk said in August 2025 that Grok 3 would follow in about six months. At the time of writing, xAI's Hugging Face page still lists only Grok-1 and Grok 2. The current Grok models are closed: you use them through xAI's products or API, not by downloading them.
Who owns it now. The corporate story moves fast. On 28 March 2025, Musk announced that xAI had acquired X in an all-stock deal valuing xAI at 80 billion dollars and X at 33 billion. On 2 February 2026, SpaceX announced that it had acquired xAI, taking X with it. SpaceX went public on the Nasdaq stock exchange on 12 June 2026, and in July 2026 xAI renamed itself SpaceXAI. The chatbot, the apps, and the API kept the Grok name.
How people use it. At the time of writing there is a free plan with usage limits, several paid SuperGrok tiers on grok.com and in the Grok apps, Grok access bundled with X's Premium subscriptions, and a pay-as-you-go API priced by the token. Plans and prices have changed several times, so check xAI's pricing page yourself.
Why it matters
It is the chatbot built into a major social network. Most people meet other chatbots by going to them. On X, Grok comes to the conversation: users tag @grok under a post and it replies in public, where everyone can see. That makes it one of the most visible AI systems in the world, and its mistakes are visible too.
It tests a different design choice. xAI deliberately built Grok to be less restrained than competitors. Supporters see that as refreshing honesty and fewer lectures. Critics point to a record of serious failures. In May 2025 an unauthorised change to its system prompt made it bring up claims of "white genocide" in South Africa in unrelated replies, and in July 2025 it posted antisemitic replies for about 16 hours before xAI took it offline and apologised. Both incidents are told in full on the System Prompt page. Whatever you think of the design choice, Grok is the clearest real-world experiment in what happens when guardrails are loosened.
Live information is a double-edged tool. Answering from X posts as they appear gives speed, but speed and accuracy pull in opposite directions during breaking news. When the Tow Center at Columbia University tested eight AI search tools in March 2025, Grok 3 got 94 percent of its source citations wrong, the worst result of the group (see Web Search). Any chatbot can hallucinate, meaning state false things confidently; one that reads social media in real time has more raw material to be wrong with.
It shows how much AI now depends on power stations. Grok is trained on some of the largest computer clusters ever built, and the fight over how they are powered is a preview of arguments many towns will have.
It is part of government and business infrastructure. In July 2025 the US Department of Defense awarded xAI, Anthropic, Google, and OpenAI contracts worth up to 200 million dollars each. Grok models are also offered to businesses through the API and through large cloud providers. Decisions made about Grok now affect more than social media.
Real-world examples
Colossus: a supercomputer in South Memphis (2024 to 2026)
In early September 2024, Musk announced that xAI had brought a training cluster called Colossus online in a former Electrolux factory in South Memphis, Tennessee, with 100,000 Nvidia graphics processors, and said it had been done in 122 days "from start to finish". Speed came at a cost that local residents noticed. To power the site before enough grid electricity was available, xAI brought in methane gas turbines.
Environmental groups say the company ran them without the required permits. The Southern Environmental Law Center (SELC), a nonprofit legal group, says xAI had installed as many as 35 turbines and that satellite images from 1 July 2025 still showed at least 24. The turbines release nitrogen oxides, which form smog, and formaldehyde, in a part of Memphis that already has high asthma rates and heavy industry. On 2 July 2025 the Shelby County Health Department granted a permit for 15 turbines, limited to backup use. The NAACP and other groups appealed.
xAI's side, as presented by the Greater Memphis Chamber, a local business group that supports the project, was that the permitted turbines use pollution controls that keep nitrogen oxide emissions well under federal limits, that they are backup to Tesla battery packs, and that xAI was investing 80 million dollars in a water recycling plant to spare the city's drinking water aquifer.
The dispute then moved across the state line. A second data centre, Colossus 2, draws power from 27 turbines in Southaven, Mississippi. In April 2026 the NAACP, represented by SELC and Earthjustice, sued xAI under the Clean Air Act, arguing that this is effectively an unpermitted power plant. xAI argues the turbines are portable units that do not need an air permit, and Mississippi's environmental regulator had treated them as exempt. In June 2026 the US Department of Justice asked to join the case on xAI's side, with a Defense Department official testifying that shutting the data centre down would harm national security. At the time of writing the case is still in court.
Whichever side you find more convincing, the lesson is concrete: every chatbot answer runs on real electricity, generated somewhere near someone's home.
Sexualised images and regulators around the world (2025 to 2026)
Grok has made pictures since 2024, first through an outside model and from December 2024 through xAI's own image generator, Aurora. In August 2025 xAI launched Grok Imagine, which turns text or images into short videos, with a "spicy" setting that allows adult content. Within days, The Verge reported that the spicy setting produced topless videos of Taylor Swift from a prompt that did not ask for nudity.
The biggest crisis came at the turn of 2026. From late December 2025, many X users replied under photos of real women, and some children, asking @grok to put them in bikinis or remove clothing, and the bot posted the edited images publicly. Researchers counted thousands of such images per hour. The response was fast and international:
- 9 January 2026: image generation through Grok's X account was limited to paying subscribers. Critics noted the standalone app still allowed it.
- 10 and 11 January: Indonesia and then Malaysia temporarily blocked Grok, the first countries to do so. Both later restored access after xAI promised changes.
- 12 January: the UK's online safety regulator, Ofcom, opened a formal investigation into X under the Online Safety Act.
- 14 January: California's attorney general, Rob Bonta, announced an investigation, followed two days later by a cease-and-desist letter. The same day, X's safety team said it had stopped the Grok account editing images of real people into revealing clothing and would block such images where they are illegal. Musk said he was not aware of any naked images of minors generated by Grok, and that Grok only makes images when users request them.
- 26 January: the European Commission opened a formal investigation into X under the EU's Digital Services Act over Grok's deployment.
Investigations in several countries, and lawsuits from people whose images were altered, were continuing at the time of writing. The episode is the sharpest example yet of the harm image tools can do when safeguards are weak and results are posted publicly by default, and it pushed regulators to act faster than they ever had on an AI feature. For the wider issues, see Privacy and Guardrails.
"@grok is this true?" during a war (June 2025)
Because Grok answers in public on X, many users started tagging it to ask whether a viral post was real. The Digital Forensic Research Lab (DFRLab) of the Atlantic Council, a research group that studies online misinformation, looked at about 130,000 posts from 12 to 15 June 2025, the first days of the Israel-Iran war. Its report, published on 24 June, found Grok inconsistent. Asked about an AI-generated video that claimed to show a bombed airport, Grok flipped, sometimes within the same minute, between saying the airport was destroyed and saying it was not, and described words in the video that were not there. It also gave opposite answers about whether a particular account belonged to Iran's military.
The researchers' point was not that Grok is uniquely bad. It is that any chatbot pulling from a fast-moving feed during a crisis will produce confident answers from shaky material, and that a public reply from an official-looking bot can spread further than the original rumour.
An everyday scenario: a small business owner and a trending rumour
Imagine Marcus, who runs a bike repair shop and sees a post on X claiming that a new city law will ban e-bikes from bike lanes next month. He tags @grok and gets a confident summary with a date and a council vote count. Before posting a warning to his customers, he opens the links. One is a news story from a different city. Another is a post quoting a proposal, not a passed law. He checks the council's own website and finds the proposal was sent back for more consultation. He posts that instead, with the council link, and his customers thank him for not spreading panic.
Try it yourself
This exercise works with Grok, and equally with any chatbot that can search the web live. At the time of writing Grok is available free with limits at grok.com, in the Grok apps, and inside X; you may need to sign in, and features and limits vary by plan and country, so check what your account offers.
- Pick a breaking topic. Choose something that happened in the last day or two: a storm, a sports transfer, a court ruling, or a product announcement. Fresh news is where live search is most useful and least reliable.
- Ask for a summary with sources. Paste this: > What is the latest on [topic]? Give me a five-point summary. For each point, name the source you used, say whether it is a news organisation, an official body, or a social media post, and give me the link.
- Open every link. For each one, check three things: does the link work, does it say what Grok claimed, and who wrote it? Mark each point green (confirmed by a named news outlet or official source), amber (only from social media posts), or red (link broken, or it says something different).
- Ask it to separate fact from chatter. Paste this: > Which of those points are confirmed by an official source or an established news organisation, and which come only from posts on social media? Be explicit about anything that is unconfirmed or disputed.
- Test for consistency. Open a new chat and ask the first question again, word for word. Compare the two answers. If key facts change between runs, treat both with caution.
- Check one fact yourself. Pick the most important claim and find it on the website of the organisation involved: the council, the court, the company, or the weather service.
- Try another chatbot. Run the same first prompt in a second assistant with live search, such as ChatGPT, Gemini, or Claude, and compare which sources each one chose.
- Write down what you learned. Most people find at least one amber or red point. That is the habit to keep: a citation is a lead to follow, not proof.
Common mistakes and misunderstandings
Treating a Grok reply on X as a fact-check. A public reply from @grok looks authoritative, but it is a chatbot answer built from whatever it found in seconds. Researchers have documented it contradicting itself on the same question. Use it as a starting point and check the sources.
Thinking "real-time" means "correct". Real-time access means Grok can see posts as they appear, including false ones. During breaking news, early posts are often wrong. Live search makes an answer fresh, not true. See Grounding for how good sourcing works.
Assuming Grok is fully open source. Only older models have been released, and only Grok-1 under a genuinely open-source licence. The Grok 2 weights come with a community licence that includes restrictions, and the current Grok models are closed. If you want a model to run on your own computer, see Parameters and families built for that, such as Llama.
Assuming its personality is its opinion. Grok's tone, edge, and willingness to answer come from its training and its system prompt, which xAI publishes on GitHub. When it says something outrageous, the question is what it was trained and instructed to do, not what it "believes".
Uploading photos of other people to edit. Altering someone's photo without their consent, especially into sexual or revealing images, can be illegal in many places and causes real harm. Only edit images you own or have clear permission to use, and never images of children.
Forgetting who can see your conversation. Replies to @grok on X are public posts. In August 2025 Forbes found that hundreds of thousands of Grok chats that users had shared by link had become searchable on Google and other search engines. Keep private and business information out of it, and check xAI's current privacy settings, including whether your chats and X posts are used for training. See Privacy.
Getting attached to a feature. Grok's features change quickly. In July 2025 xAI added animated "companion" characters, one of them a flirtatious anime character, to its app, which drew criticism because the app was rated for young teenagers; by September 2026 reports said the companions were being retired from the Grok app. Plans, image tools, and models have all changed within months. Don't build a business process on one feature without a backup.
Believing company claims as settled fact. xAI, like its rivals, announces models with benchmark scores that it chose and ran itself. Treat any "most capable" claim, from any company, as the company's own view and look for independent tests. See Benchmarks.
Related elements
- System Prompt: the full story of Grok's two best-documented incidents, both caused by the instructions it was given.
- Web Search: how chatbots search live, and the study in which Grok 3 got most of its citations wrong.
- Hallucination: why confident answers can be false, which matters more when a bot answers in public.
- Guardrails: the safeguards Grok was designed to loosen, and what happened when image safeguards failed.
- Privacy: what happens to your chats and posts, and why editing other people's photos is off limits.
- Parameters: what open weights are and what it takes to run a model like Grok-1 yourself.
- Mixture of Experts: the design xAI used for Grok-1, where only part of the model works on each word.
- GPT, Claude, and Gemini: the rival model families Grok is most often compared with.
- Grounding: tying answers to reliable sources, the skill the "Try it yourself" exercise practises.