Grok has the oddest reputation of any major chatbot. Ask a developer about it and you’ll hear about benchmark scores. Ask anyone who has spent a week on X and you’ll hear about the time it insulted them, correctly, in three words. Both things are true, and neither is a great guide to whether it’s worth your time.
So here’s the practical version.
What Grok actually is
Grok is the assistant built by xAI, Elon Musk’s AI company, and wired straight into X. It launched for Premium+ subscribers in late 2023 and has been rebuilt roughly twice a year since: Grok 2 in August 2024, Grok 3 in February 2025, Grok 4 that July. There’s now a standalone app, a web interface, and an API.
The training runs scaled hard. Grok 3 was trained on a Memphis cluster that grew from 100,000 Nvidia H100 GPUs to roughly 200,000, using about ten times the compute of its predecessor.
What separates Grok in daily use isn’t raw capability. It’s the plumbing. Grok reads X posts as they happen, so it can answer questions about events that haven’t reached a training set or a news article yet.
The live feed is the real differentiator
Ask ChatGPT about a protest that started six hours ago and you’ll get a careful answer about a knowledge cutoff. Ask Grok and it pulls posts from people on the ground, quotes them, and tells you which accounts said what. Whether those accounts are right is a separate problem, but you get the raw material fast.
The genuinely useful cases:
- Breaking stories, where the first twelve hours of a major event are often covered faster on X than anywhere else
- Sentiment checks, like how a product launch or a cup final actually landed with a specific audience
- Finding the original source of a claim that’s been screenshotted and reshared a hundred times
- Tracking niche communities through a crisis, a patch, or a collapse
There’s a bias baked in. X is not the world. It’s a few hundred million accounts skewed toward tech, politics, sports, and English. Grok will tell you what X thinks, which is not the same as what is true or what most people think.
Where Grok pulls ahead
Reasoning, if you pick the right mode
Grok 4 shipped with a think mode that works a problem before answering. On Humanity’s Last Exam, a deliberately brutal graduate-level benchmark, Grok 4 Heavy scored around 25%, the top published number at launch, against roughly 21% for standard Grok 4. Those figures sound low. The test is designed to be near-impossible, and the relative gap is what matters.
In practice, think mode is better at multi-step maths, debugging, and anything where the first plausible answer is wrong. It’s also slower and burns through quota faster.
Tone that doesn’t sound like a compliance department
Grok argues. It will tell you an idea is bad. Some people find that refreshing after years of replies opening with “Certainly!”. Others find it exhausting by week two. The personality is a deliberate product choice, and it’s why a lot of users stay.
Zero friction inside X
If you already live on the timeline, highlight a post, hit the Grok button, and get a summary, a translation, or a fact-check without switching tabs.
How it stacks up against ChatGPT and Claude
For general writing, Claude is still the smoother read, and ChatGPT’s ecosystem of connectors and custom tools is deeper. Grok wins on live X data, a looser voice, and tight integration if you’re an X power user. On coding, all three are close enough that benchmark gaps rarely survive contact with a real codebase.
No assistant wins everywhere, which is why this list of ChatGPT alternatives that beat it at something specific is worth a read before you commit to another subscription.
Where it still trips up
Confident nonsense. Grok hallucinates at rates comparable to its peers, and its confident tone makes the errors harder to spot. It has invented court cases and mangled statistics with total composure. A detailed breakdown of what xAI’s chatbot does better than anyone else, and where it keeps tripping, covers the pattern properly.
Moderation whiplash. xAI’s guardrails are loose by design, and it has backfired loudly. An update in July 2025 that let the Grok account post on its own produced a run of antisemitic content within hours before being pulled. Earlier that year it answered a question about a political controversy by treating “white genocide” rhetoric as fact. Regulators in Poland, Turkey, and elsewhere have noticed.
Thin sourcing. When Grok quotes posts it doesn’t always weigh them. A single viral claim can read as solid as a wire report.
Images, video, and a filter set to “off”
Grok generates images through Aurora, built on Black Forest Labs’ Flux models. Quality is genuinely high: photoreal portraits, decent hands, prompt adherence that beats most free generators. Moderation is the problem. Within days of launch, users were producing copyrighted characters and real politicians in compromising scenes. xAI tightened, loosened, and landed somewhere unpredictable.
Prompting works like any diffusion model: describe the shot, not the idea. This step-by-step FLUX workflow for turning rough ideas into finished images maps almost directly onto Aurora, since it’s the same model family. The prompt structure behind ChatGPT images that actually work transfers cleanly too: subject, framing, light, lens, style, then detail.
Grok Imagine, the video spin-off, arrived in August 2025. It’s fast, strange, and not ready for client work.
What it costs
There’s a free tier on X and in the app, capped every couple of hours. Enough for casual questions, not enough for an afternoon in think mode. SuperGrok runs about $30 a month and lifts the caps. X Premium+, around $40, bundles Grok with the rest of X’s paid features. SuperGrok Heavy at $300 unlocks Grok 4 Heavy and much higher limits. The API is priced per million tokens, in line with everything else on the market.
How to get better answers out of it
- Pick the mode deliberately. Think mode for maths, code, and logic. The fast model for summaries and rewrites.
- Ask for sources. “Cite the posts you’re drawing on” makes bad sourcing visible immediately.
- Paste the material. Drop in the article or thread and tell it to work only from that.
- Push back once. Grok holds its position until challenged, then over-corrects. A single “are you sure?” produces a much more careful second answer.
The questions Grok hasn’t answered yet
Two things will decide whether it becomes more than a niche tool for X users.
The first is trust. A chatbot that can be talked into producing hate speech, or that repeats a viral claim as fact, is hard to put in front of customers or students. xAI’s instinct has been to loosen the rules and patch afterwards, which keeps the product interesting and the headlines coming. Enterprises don’t buy interesting.
The second is whether the X integration stays an advantage while the platform’s ad revenue wobbles. Grok’s live-data edge exists because X is still the fastest place a lot of news breaks. If that stops being true, the edge disappears quietly.
For now, the sensible play is narrow. Keep Grok for the jobs it genuinely does better, like breaking news, sentiment, and a fast read on a noisy crowd, and keep a second assistant around for the work that has to be right the first time.