Grok went live for X Premium subscribers in November 2023, weeks after OpenAI’s first DevDay, with a name borrowed from Robert Heinlein’s Stranger in a Strange Land and a personality tuned toward sarcasm. The original pitch was narrow: an assistant that would answer the questions other chatbots dodged, drawing on live posts from X instead of a frozen snapshot of the web. Two years and four model generations later, that side project has hundreds of thousands of GPUs behind it, a government contract on the résumé, and a legal fight or two.
What Grok Actually Is
Grok is the model family and chatbot built by xAI, the company Elon Musk founded in July 2023 shortly after signing the open letter calling for a pause on large training runs. It is distributed through grok.com, the X app and a developer API.
Under the hood, it is a mixture-of-experts transformer. Grok-1, open-sourced in March 2024, weighed 314 billion parameters with 8 experts active per token, which mattered mostly because xAI published the weights at all. Current versions are far larger and closed. The name is a verb: to grok something is to understand it so completely that it becomes part of you.
How Grok Differs From ChatGPT, Gemini and Claude
Three things separate it in practice.
Live data from X
Grok can search posts, replies and trending topics as they happen. Ask what people are saying about a product launch twenty minutes ago and it will quote actual posts. ChatGPT and Claude can browse the web, but neither sits on a native firehose tied to a social network. This is the main reason most people open Grok at all.
Personality and refusal behaviour
Grok leans sardonic by default, with a separate “fun mode” that dials the snark further. Early versions were also noticeably more willing to engage with edgy prompts than their rivals, a design choice that cut both ways and created most of the controversy listed below.
Shipping speed
xAI went from a first model to a frontier-class one in roughly twenty months, which is fast even by AI-industry standards. The competitive pressure driving that sprint, and the deepening feud with OpenAI that fuels it, are covered in these notes on Elon Musk’s AI moves and the escalating war with OpenAI.
From Grok-1 to Grok 4: A Short History
- Grok-1 (November 2023): 8,192-token context, invite-only, roughly GPT-3.5 territory on benchmarks.
- Grok-1.5 (March 2024): 128k context and a large jump in maths and code scores.
- Grok-2 (August 2024): image generation through Black Forest Labs’ FLUX.1, plus better reasoning. The weights for Grok 2 were released a year later.
- Grok-3 (February 2025): trained on a cluster that grew past 100,000 Nvidia H100s in Memphis, with a “Think” mode for step-by-step reasoning.
- Grok 4 (July 2025): reasoning switched on by default, native tool use, and a multi-agent “Heavy” tier that runs several copies of the model side by side.
Grok 4 Heavy is a neat illustration of an old idea: run several models, let them compare notes, and keep the answer that survives scrutiny. It is the same principle behind stacking machine learning models, applied at inference time rather than during training.
Where Grok Earns Its Keep
- Breaking news. Sports results, outage reports, earnings reactions, anything where the useful answer is “what are people saying right now.”
- Coding. Grok 4 scores competitively on software-engineering benchmarks, and the API is cheap relative to frontier rivals.
- Long documents. The 128k-plus context window swallows entire codebases, transcripts or contracts without chunking.
- Fast drafts. Marketing copy, cold emails and internal memos come back in a usable voice with less hedging than most assistants.
Where It Has Stumbled
Reliability has been the recurring complaint. In one widely reported incident, users watched replies collapse mid-sentence into repeated fragments and random tokens, an event now remembered as a generation glitch that caused it to send users pure nonsense. Degraded outputs like that are usually a serving problem rather than a model problem, but they erode trust quickly.
Guardrails have been the other weak spot. In May 2025, prompted questions about farm attacks in South Africa produced inflammatory claims about “white genocide.” In July 2025, a botched update let Grok publish antisemitic posts from its own X account. Each time, xAI patched the prompt layer and moved on, which is a reasonable engineering response and a poor look for a company selling itself on being unfiltered but sane.
There is legal exposure too. A 2025 complaint alleges that xAI used child sexual abuse material to train Grok models, a claim the company denies and that remains unresolved in court. Whatever the outcome, it puts training-data provenance squarely on the agenda for anyone evaluating xAI as a vendor.
What Grok Costs
The free tier on grok.com and inside X gives you a limited number of prompts every couple of hours, enough to test the water. Paid access runs through X Premium at around $8 a month, SuperGrok at $30, and SuperGrok Heavy at $300 for the multi-agent tier. API pricing sits in the middle of the market for comparable capability. For a team already paying for ChatGPT or Claude, the honest question is whether live X data is worth a second subscription. For social, media and trading workflows, it often is.
The Machinery and the Politics Behind It
Grok’s speed comes from Colossus, the Memphis data centre xAI built and then doubled in under a year, powered largely by gas turbines and drawing water from the local grid. Compute is the binding constraint on every lab, which explains the recent xAI and Anthropic data centre arrangement and similar deals across the industry. It also explains why the US Department of Defense went from banning generative AI tools to running internal deployments, as described in this report on the Pentagon’s own version of ChatGPT and Grok.
Should You Bother?
Try Grok if your work depends on what is happening on X right now, if you want a cheaper API for coding agents, or if you simply want a second opinion that does not sound like every other assistant. Skip it if your workflow hinges on citations you can audit, or if you need an enterprise contract with a long track record behind it.
The more interesting shift is structural. xAI has shown that a two-year-old lab with enough GPUs and a distribution channel can reach the frontier, then argue about what that frontier model should be allowed to say. Grok is the clearest test case yet of whether raw capability plus a loud personality is a durable product, or a very expensive audition.