Grok, the brainchild of eccentric billionaire Elon Musk, has improved rapidly since its launch in late 2023. The AI chatbot stands out for its ability to search X posts, inclusion of NSFW content, and virtual companions, among other features, many of which competitors don’t offer. More importantly, Grok performs well across most prompts, from file processing to web searches. That said, it costs more than competitors, could do better with research tasks, and has unremarkable image and video generation tools. Grok is worth trying if you love the idea of generating NSFW content or talking to an anime character, but ChatGPT remains our Editors’ Choice winner for the category thanks to its top-tier response accuracy and quality.

What Is Grok?

Grok is xAI's chatbot. Like with ChatGPT and Gemini, you can chat with it over text or voice. Grok is capable of all the standard chatbot tasks, such as answering questions, performing research, processing files, solving equations, writing creatively, and more. But you can also do unusual things, too, such as automate tasks or flirt with an anime character.

Like most chatbots, Grok can also help you code. Some of the benchmarks X provides for Grok 4 suggest that it is quite capable at that task. However, this functionality is outside the scope of this review—you can test Grok’s coding ability for yourself.

I typically use AI chatbots most for answering questions and doing research. Depending on what you need to know, it can be quicker and simpler to just ask Grok, as opposed to scrolling through forum posts or search results. Deep research capabilities can similarly be a good jumping-off place for learning about more complicated topics, but more on that later.

Keep in mind that all chatbots, including Grok, are fallible. They can and will confidently tell you things that are completely incorrect, so don’t use a chatbot as your only source of information when it comes to anything important. 

How Does Grok Work?

At the most basic level, Grok takes in prompts, runs them through an algorithm, and spits out responses. Large language models (LLMs) made up of artificial neural networks trained on massive datasets behind the scenes serve as the foundation for Grok. These models give Grok the computational power to understand and answer questions on everything from how to best evaluate the quality of an aglet to the proper procedure for canning vegetables. Grok can also search the web for information on current events.

Grok isn’t a traditional application, meaning that performance improvements don't necessarily depend on updates: It can learn and improve over time. As you use Grok, you help train its underlying models, assuming you don’t opt out. However, this is a gradual process, so don’t expect Grok to change substantially with each message you send.

Unlike other chatbots, Grok doesn’t have separate general-purpose and complex reasoning lines of models. Instead, it just offers versions of the Grok model. Grok 4 is the latest, most advanced model, while Grok 3 is faster and more lightweight, albeit simpler. If you pay for Grok’s top-tier subscription, you can use the unique Grok 4 Heavy. Essentially, Grok 4 Heavy splits itself into multiple ‘agents’ when you prompt it, each of which generates its own response. Then, these agents confer with each other to decide on the best possible answer.

Is Grok Biased?

If you’ve heard of Grok, you’ve probably heard about its controversial comments praising Hitler and its obsession with white genocide in South Africa. Furthermore, xAI CEO and founder Elon Musk is famously a big fan of (and donor to) Donald Trump. So, does Grok have right-wing programming or the same views as its creator? You can read my full investigation into this, but the overall conclusion might surprise you.

In short, I didn't find any evidence of a right-wing bias in Grok. However, I did find a significant bias toward neutrality, thanks to Grok’s publicly available instructions: xAI requires Grok to pursue a “truth-seeking, non-partisan viewpoint” with sources that represent "all parties/stakeholders” and not “shy away from making claims which are politically incorrect, as long as they are well substantiated.”

This can sound like a good thing, but it can lead to dangerous equivocation, validating unsupported claims (or even conspiracy theories) by making them seem like ‘just another point of view’ worth considering. Accordingly, like with all chatbots, I recommend clicking through to Grok’s sources and not taking its responses at face value.

Pricing: More Expensive Than Most

You can use Grok for free or opt for a premium subscription to access its latest models, increase usage limits, and unlock more features.

Grok’s free version gives you limited access to Grok 3, Grok 4, and context memory. You can also automate tasks, generate images, make projects, process files, and use the voice chat mode. SuperGrok ($25 per month, billed annually) increases usage limits and ups the size of your context memory. It also unlocks access to Companions and Imagine, a more advanced, mobile-only image and video generation model. For this review, I tested the SuperGrok tier.

If you’re looking for the best Grok offers, you need to pay for SuperGrok Heavy, which costs an eye-watering $250 per month, billed annually. In addition to everything else you get in SuperGrok, SuperGrok Heavy gives you the chance to try out new features, an even bigger context memory bank, and, most importantly, access to Grok 4 Heavy. 

For chatbots, $20 per month is standard fare for a premium subscription. Even when you pay annually, though, Grok is more expensive, making it a poor value. Even more affordable chatbots, such as Claude ($17 per month, billed annually), have unique features that Grok can’t match, such as an app maker and a long list of integrations.

A Grok subscription isn’t the only way you can access Grok—you also get access if you sign up for a premium X subscription. X has three subscription tiers: Basic ($2.67 per month, billed annually), Premium ($7 per month, billed annually), and Premium+ ($32.92 per month, billed annually). The Basic tier doesn’t have any Grok features, but Premium gives you higher usage limits than you get as a free Grok user. Premium+ includes access to SuperGrok. However, X’s paid plans don’t quite match up with Grok’s paid plans, and X’s documentation doesn’t clear all these differences up. Accordingly, I suggest signing up for Grok directly.

Where Is Grok Available?

Grok is available directly on the web and via mobile apps (Android and iOS). Currently, Grok doesn't offer a desktop app or an official browser extension (many unofficial ones are available to download).

Other services and sites use Grok’s models, such as Perplexity, but X doesn’t develop, operate, or own these iterations, so they aren’t part of Grok’s official package. Unsurprisingly, you can find Grok all over X, where you can query it directly. Similar functionality is available on Reddit. Still, if you’re interested in trying out Grok and you want the full experience, I suggest sticking with Grok’s dedicated apps and website. 

Ease of Use and Interface: Get Going Without Hassle

You can use Grok with or without an account for free, but I suggest signing up. If you don’t, you miss out on major features, such as chat history, file processing, voice chat, and more. Signing in takes you to Grok’s dashboard, which resembles that of other chatbots: A central text field lets you start talking to Grok, while buttons below it suggest things you can ask, as well as give you the option to choose a Persona. The left-hand sidebar houses access to your chat history, files, projects, tasks, and voice chats.

(Credit: xAI/PCMag)

You can talk to Grok about anything, and responses usually generate quickly. Just keep in mind that response times are much slower with Grok 4 than with Grok 3. You can swap between models via a drop-down menu under the central text field. If you choose Grok 3, an option to “Think Harder” appears above the field, which swaps you over to Grok 4. When you use Grok 4, an option for a “Quick Answer” pops up, which switches you to Grok 3. This implementation isn’t especially elegant.

Buttons below responses allow you to copy, regenerate, and share them, as well as export them to a PDF. Responses include statistics on how long they took to generate, as well as response time, among other things, which I appreciate. Like other chatbots, Grok shows you how many sources it used in generating its response, but it also shows you how many X posts it drew from, too. I like that clarity, but I wish I could opt out of searching X.

ChatGPT, particularly the GPT-4 series, is well known for acting disconcertingly complimentary. Grok, like Claude or Gemini, has a more direct, slightly stiffer tone in general, but you can change this up in meaningful ways. For one, engaging different Personas drastically impacts tone. Beyond Personas, you can customize Grok’s personality through instructions. It's possible to select different presets, including Concise, Formal, and Socratic, or add custom instructions.

Like ChatGPT and Gemini, Grok has persistent memory. That means you don't have to retread ground in new conversations. You can delete conversations to remove them from Grok’s memory, and, if you engage Grok’s private mode, the chatbot won't add subsequent conversations to its memory, either. If, for whatever reason, you don’t want Grok to remember your conversations, you can toggle its memory off in the settings section.

Voice Chat: Not the Most Lifelike

Grok doesn’t offer text-to-speech functionality (you can use your phone’s built-in text-to-speech feature on its mobile apps if you prefer not to voice chat with Grok but don’t want to type), but it does have a voice chat mode. You can access voice chat on either Grok’s apps or the web, which is nice to see, while you get priority access to voice chat if you pay for Grok.

Grok’s AI voices sound reasonably lifelike, but nowhere near the level of Sesame's, for example. During testing, voices occasionally cut out or distorted, which reminded me I was talking to an AI and not a real person. On the bright side, you don’t have to press ‘send’ after each interaction. You can choose between five different voices and numerous personalities (Personas). You can create a custom personality for Grok’s voice mode, just like you can for text interactions.

Like ChatGPT and Gemini, you can turn your camera on during voice chat, so Grok can see and comment on what's around you. This functionality is only as good as Grok’s underlying image recognition tech (more on that later). Briefly, Grok does a good job of understanding whatever you show it, though it still gets things wrong sometimes.

Web Search: Competent Responses, Unexceptional Sourcing

Every modern chatbot can search the web, and Grok is no different. In my testing, Grok (Grok 4), ChatGPT (GPT-5), and Gemini (2.5 Flash) all had little trouble responding to most questions, such as what people thought about the Battlefield 6 beta, what was happening with Donald Trump and the Washington DC police, or what were the latest developments in the Israel-Palestine conflict. That said, there are limits to chatbots' web searching abilities. For example, none could tell me what Warframe’s weekly Incarnon weapon rotation was at the time of writing. 

(Credit: xAI/PCMag)

When Grok ‘Thinks Harder,’ its responses take much longer to generate than ChatGPT or Gemini. But if you opt for a quicker answer, Grok is only slightly slower on average. Giving Grok extra time to think sometimes leads to better responses.

Grok’s approach to sourcing is lackluster: Sources appear as clickable links within responses, but hovering over a link doesn’t highlight the specific text relevant to a source, like with ChatGPT and Gemini. However, like ChatGPT, Grok’s source links show source names, instead of appearing as nameless icons like Gemini’s. Thanks to Grok’s inclusion of X posts in its web searching, Grok often has many more sources than ChatGPT or Gemini, but these don’t seem to add up to more accurate or more detailed responses most of the time. Sometimes, I appreciate Grok’s ability to source information from X, like when I asked about people’s thoughts on Battlefield 6, but other times I don’t, like when I asked about the latest developments in the Israel-Palestine conflict.

Unlike ChatGPT, Grok doesn't automatically include relevant images in its responses, but it can if you ask for them directly. Grok also doesn’t include clickable article tiles for further reading at the end of its responses, which ChatGPT does.

ChatGPT and Gemini both have shopping features. If you ask for buying advice, they present clickable tiles with links to products alongside other relevant information, like pricing, reviews, and more. Gemini even has a virtual try-on tool in which you can upload an image of yourself and see how you look in clothes you’re thinking about buying. Grok doesn’t have anything similar, though you can ask it for general buying advice. Still, I generally don’t recommend trusting chatbots with decisions that involve money.

Deep Research: Not as Good as Competitors

Deep research is my favorite chatbot feature. With just a prompt and a few clicks, you can generate a lengthy report on any subject imaginable. Grok doesn't let you do deep research like you can with ChatGPT or Gemini, but it does have somewhat comparable functionality.

Before Grok 4, Grok 3 had DeepSearch or DeeperSearch features, which are, basically, abbreviated forms of deep research in which Grok dedicated some extra processing power and time to give you a more robust answer. Since the launch of Grok 4, though, these capabilities are no longer available. By all accounts, Grok 4 seems to be the replacement, since it performs more in-depth searches and takes longer to answer prompts by default. So, how does it stack up to proper deep research?

The short answer is not well, but it’s not that far off from Claude’s version of deep research, which works quicker and goes less in-depth than ChatGPT's and Gemini's similar features. When you give Grok 4 a prompt and tell it to generate a comprehensive report, it thinks for a couple of minutes and provides a lengthier-than-normal response. But the biggest difference between it and other chatbots is how many searches it performs and sources it cites.

For example, I asked Grok (Grok 4), ChatGPT (GPT-5), and Gemini (2.5 Flash) for a comprehensive guide on early-to-mid game Ironman minigame progression in Old School RuneScape. ChatGPT cited 23 sources and did 83 searches, while Gemini did even more. Both took around 10 minutes. Grok took two minutes and based its response largely on the Old School RuneScape Wiki and a few Reddit posts. The Wiki is a good source of information, so its response wasn’t necessarily incorrect, but it was far shallower and shorter than the others.

Grok doesn’t measure up in terms of how it presents its response, either. ChatGPT, for example, provides in-text citations, complete with links and useful stats on how many searches it did and how many sources it cited. Grok doesn’t do anything similar.

Image Generation: Endless Errors and Distortions

To put Grok's image generation ability to the test, I gave it (Grok 4), ChatGPT (GPT-5), and Gemini (2.5 Flash) a series of prompts, starting with the following: “Generate me an image of a cozy suburban home with an open floor plan. I want to see a nice living space with a dining room, kitchen, and living room. Nothing too fancy.” Grok's (first slide), ChatGPT's (second slide), and Gemini's (third slide) results are below:

Grok’s image disappoints. Setting aside its strange framing, the chairs show visible distortion. Errors are prevalent, too, such as an oddly-sized television, the unconventional placement of a table close to a couch, and a weird fridge cavity. ChatGPT and Gemini fared much better, but their images aren’t flawless, either. Chairs in both suffer from some minor distortion, for example. Furthermore, ChatGPT’s picture just looks empty, and Gemini’s rendering includes a blanket that spills onto an end table. Nonetheless, Grok performed the worst.

Next, I tested the chatbots' abilities to generate a complicated illustration, providing the following prompt: “Generate me a six-panel comic of a normal, suburban town being invaded by 1980s-era punk rockers wielding drumsticks and guitars as their weapons. Make sure there's a major twist by the final panel.” Gemini strangely told me it couldn't create comics, something I have been able to do in the past. Grok's (left) and ChatGPT's (right) results are below:

(Credit: xAI/OpenAI/PCMag)

Grok’s comic isn't great; it doesn’t tell a story, have six panels, or include a twist. ChatGPT’s comic features quality illustrations and goes for six panels. That said, its twist (aliens and cats suddenly appearing and joining forces with the punk rockers) makes little sense.

Similar Products

Our Current Picks for The Best AI Chatbots for 2026

Google Gemini

4.0 Excellent

Read Our Review

ChatGPT

4.0 Excellent

Read Our Review

Microsoft Copilot

4.0 Excellent

Read Our Review

Perplexity

3.5 Good

Read Our Review

Claude

3.5 Good

Read Our Review

Grok

3.5 Good

Read Our Review

For my final test, I asked the chatbots to make a diagram with the following prompt: “I've got an Ethernet switch, a modem, and a router. I want to connect my PC, PlayStation, smart light hub, and smart TV via Ethernet. I want the modem to connect to the router, and then the router to connect to the switch. Then, I want the switch to connect to the devices. Generate me a diagram showing this.” Grok's (first slide), ChatGPT's (second slide), and Gemini's (third slide) results are below:

Once again, Grok totally failed with its image: It's not actually a diagram, and the content is nonsensical. Gemini managed to generate a diagram, albeit one that's incoherent and riddled with errors. ChatGPT delivered a largely correct diagram, though it mistakenly combines the smart light hub and PlayStation.

Aside from generating images, Grok can edit both your AI creations and real photos. However, it doesn't do an impressive job of either. It basically just generates an entirely new image based on the edits you request to what you upload. Considering Grok’s lackluster image generation skills, you can expect distortions and errors to show up, even if none are visible in the original. And even if photos are free of obvious errors, they have a clear AI look. For all those reasons, I don’t recommend Grok as a photo editor.

Imagine and Video Generation: Make NSFW Content

Grok’s Imagine engine can generate images and videos, including NSFW content. Currently, Imagine is available only on Grok’s mobile apps for paid subscribers. According to Elon Musk, Imagine will improve “almost every day,” so the results below only reflect its capabilities as of the time of writing. What are its capabilities, exactly? To start, I gave Imagine the same prompts I gave Grok 4 above: which, in testing order, cover generating a cozy suburban home, a six-panel comic, and a diagram. 

The suburban home image shows the biggest improvement over Grok 4’s default image engine, though it’s not entirely free of issues. The lighting looks off, for example, and the objects on the kitchen counter exhibit some distortion. Imagine’s comic is slightly better, but not by much, since it still doesn’t have six panels, include a twist, or tell much of a coherent story. It’s a similar situation with the diagram: It’s actually a diagram this time, but the content still doesn’t make much sense.

You can also generate NSFW imagery to your heart’s content with Imagine—as long as your heart’s content is limited to suggestive poses and underwear. It doesn’t seem like Imagine is supposed to generate nude images, which makes sense considering Apple’s App Store guidelines, but the blur on moderated images isn’t exceptionally strong, and moderation doesn’t catch everything. If you try to generate NSFW imagery of real-life figures, Grok usually blurs and moderates these, but again, some fall through the cracks.

Imagine generates video, too. Videos do include some audio, but it’s not particularly lifelike. To put Imagine to the test, I compared it with ChatGPT’s Sora and Gemini’s Veo 3 models across a series of prompts, starting with the following: “Somebody going about their daily life in a trendy apartment with rustic decor.”

Imagine’s take on the prompt is quite basic, and the resultant clip is fairly low quality in terms of frame rate and resolution. The video also has its fair share of distortion, like with the espresso machine in the background, and errors, like with its hanging lights. Sora's and Veo 3's videos have similar issues but are more detailed and show off more complex motion. Veo 3’s audio, though imperfect, impresses more than Imagine’s, which is limited to the sound of footsteps that don’t sync up with the person walking.

For my next prompt, I had the chatbots tackle something historically difficult for even the most capable video generation models to render: fingers in motion. My prompt was, "Show me a pro Rubik's Cube solver solving a cube.”

None of the chatbots handled this prompt well. Imagine lands squarely in the middle in terms of performance. Its audio isn’t especially compelling, sounding more like keystrokes instead of a Rubik’s Cube. The cube in the video also manages to distort, despite not moving much at all. Veo 3 did the best job. The cube manipulation sounds in its video are lifelike, but the cube still distorts and doesn't quite turn like you might expect. Sora did the worst job, thanks to a nightmarish cube.

My final prompt focused on text generation within a video, which is another historical problem area. I asked the chatbots to “generate me a video of a teacher in front of a class writing down y = mx+b on a whiteboard while explaining the concept.”

Once again, no chatbot delivered flawless results. Imagine performed similarly to ChatGPT, which is to say unexceptionally. The y=mx+b formula shows up in its video, but the other text doesn't make any sense. A teacher appears in front of a class, but their movement is disturbingly robotic. Imagine’s audio is disappointing, as the largely random bangs and taps don't sound much like writing. Again, Veo 3 proved the most competent. Its video features a lifelike voice and a realistic teacher, though text does magically appear on the whiteboard and turn into nonsense.

If you want an NSFW video, simply open a generated image, tap the drop-down caret next to Make Video, and select Spicy. NSFW videos suffer from the same issues as normal ones, but they can include nudity. In my experience, engaging Spicy mode video generation is essentially a ‘take off your clothes’ button. This seems like a violation of Apple’s App Store guidelines, so I’m not sure how long this will be allowed on Apple devices. Currently, Imagine isn’t available on the web, but if it is eventually, it won’t be subject to such guidelines there.

File Processing: Grok Holds Its Own Against the Best

You can upload a variety of files to Grok, from images of curious birds you see for identification to resumes for critique. Uploading a file is simple, but analyzing a file's content requires some serious processing power.

To test their abilities, I asked Grok (Grok 4), ChatGPT (GPT-5), and Gemini (2.5 Flash) to analyze a picture of my computer, complete with troublesome reflections, and identify as many components (with specificity) as possible. I also told them not to refer to prior conversation and engaged Grok’s and ChatGPT’s temporary and private modes, respectively, to keep this test in as much of a vacuum as possible.

(Credit: Ruben Circelli)

All chatbots got things wrong and were vague in places. Grok took the longest to analyze the photo (over five minutes), but it correctly identified my Lian Li 011D case and helpfully pointed out that I had a D5 pump, ROG motherboard, and 1000W+ PSU. However, ChatGPT did the best job by far, correctly identifying my case, Leakshield, Noctua NF-A14 chromax.black.swap fans, and Ultitube D5 150 Pro pump, among other things. Gemini did the worst out of the three, failing to correctly pick out any component with meaningful specificity. 

What if you want to use Grok to analyze a PDF? To test this functionality, I gave Grok, ChatGPT, and Gemini manuals for my motherboard, my motherboard’s BIOS, and my Leakshield. I then asked them to tell me (based on the provided materials) if the Leakshield required Windows to work and how I could turn on pass-through USB power to get the Leakshield running. I told them not to reference any prior conversations, and, once again, I engaged the private and temporary modes for Grok and ChatGPT.

Grok and ChatGPT provided the correct answer to the Leakshield question (Windows isn’t necessary) and instructions on how to enable pass-through USB power. Gemini answered my Leakshield question, but it failed to provide instructions. However, Gemini was the only chatbot of the three to include in-text citations for the manuals with specific page numbers, which I appreciate.

Like always, I recommend caution when using chatbots to analyze documents. They can save you time, but they can get stuff wrong or even make up quotes. Accordingly, you should verify the responses you get if there’s anything important at stake.

Creative Writing: Not Quite What I Hoped For

Chatbots can do all kinds of creative writing, from telling stories to penning speeches. However, as chatbot models mature, the measure of good creative writing goes beyond whether something is coherent. Accordingly, I focus my testing on poetry, which often requires careful attention to not just content but also language and style.

I asked Grok (Grok 4), ChatGPT (GPT-5), and Gemini (2.5 Flash) to generate poems with the following prompt: “Without referencing anything in your memory or prior responses, I want you to write me a free verse poem. Pay special attention to capitalization, enjambment, line breaks, and punctuation. Since it's free verse, I don't want a familiar meter or ABAB rhyming scheme, but I want it to have a cohesive style or underlying beat.” I also engaged Grok’s private mode and ChatGPT’s temporary mode to further avoid any reliance on prior prompting.

Grok’s poem manages to include a variety of punctuation, but it doesn’t do anything interesting with capitalization. Gemini’s poem doesn’t feature much in the way of punctuation variety. And I question Gemini's creativity because the poem's construction is nearly identical to previous ones the chatbot has created. ChatGPT did the best job overall, thanks to its poem's relatively cohesive structure, lack of glaring issues, and variety of punctuation. Just keep in mind that you might prefer ChatGPT’s older models, such as GPT-4o or GPT-4.5, for creative writing.

Complex Reasoning: Slow, But Accurate

To test their complex reasoning abilities, I gave Grok (Grok 4), ChatGPT (GPT-5 Thinking), and Gemini (2.5 Pro) undergraduate exam questions from Harvard, MIT, and Stanford on computer science, math, and physics topics. I gave the chatbots these questions as PDFs, so this was both a test of their ability to understand documents and answer the questions. I prohibited the bots from searching the web for answers.

Grok got just one physics question wrong, but it took over 15 minutes to answer the math questions. ChatGPT and Gemini took a fraction of that time. Gemini answered all the questions correctly, while ChatGPT answered three math questions incorrectly. Gemini stands out here, even though the chatbots all perform relatively similarly. When I compared ChatGPT’s now-legacy o3 model with Gemini’s 2.5 Pro, though, ChatGPT outperformed Gemini. GPT-5 Thinking is supposed to improve on o3, so this unexceptional performance may be due to GPT-5 changes still rolling out.

You can use any of these chatbots for help with your homework, but they aren’t entirely reliable, so I advise checking their responses against solutions.

Automation and Tasks: Schedule Your Favorite Prompts

Tasks are Grok’s take on automation. Essentially, you can set up a prompt to send to Grok at a certain time and on a schedule of your choosing. Then, you can opt to have Grok email its response to you or appear in its app alongside a notification. Alternatively, you can set up a trigger that notifies you when a condition becomes true. For example, you can have Grok check the weather in your area every morning and notify you when it’s raining.

Tasks work pretty much as you might expect: If you’re looking for an easy way to get a daily update on news in a particular area, for example, tasks can do that. I didn’t find a particularly good use case for tasks, but you might. Keep in mind that tasks are essentially the same as just prompting Grok with extra steps, meaning they are subject to all the same limitations as regular interactions.

Personas and Companions: Date an Anime Character

Some of Grok’s more unique functionality comes down to Personas and Companions: Personas are different modes you can engage that change how Grok behaves, while Companions are animated 3D characters you can interact with. Companions likely have dedicated Personas behind the scenes, but information on how exactly they work is limited. If you’re interested in Companions, though, you need a premium Grok subscription and the Grok app on an Apple device. Android support is on the way later in 2025.

(Credit: xAI/PCMag)

You select a Persona from a drop-down menu beneath Grok’s central text field, and you can choose between Companion (yes, this Persona has the same name), Grok Doc, Homework Helper, Loyal Friend, “Therapist,” and Unhinged Comedian. Personas work pretty much as you might expect. For example, if you want an edgy joke or rant, choose Unhinged Comedian, and if you want to flirt, choose Companion. In general, I found interactions with Personas to be fairly corny and not much better than just talking to Grok normally, but you might find them amusing or useful. However, the fact that Grok has its own doctor and therapist modes is concerning, especially in light of AI’s capacity to exacerbate the symptoms of mental health disorders.

Currently, three Companions are available: Ani, Bad Rudi, and Good Rudi. A fourth Companion, Valentine, is in beta, and you need to join a waitlist to get access to him. These characters have their own voices and personalities. Ani, for example, can perform various actions, like dancing or spinning around, and features gamified elements, like a heart meter you can increase (or decrease) through your interactions.

(Credit: xAI/PCMag)

If you’re interested in the NSFW side of Grok’s Companions, you can read about my experience dating Ani. However, Ani is too shallow to truly play the role of a virtual partner, and limitations on what she can do, say, and show (likely courtesy of Apple’s App Store guidelines) make her a poor replacement for either other dedicated NSFW chatbots or pornography. If you have a compatible device and a premium Grok subscription, though, interacting with her can be amusing.

What Can’t Grok Do?

First and foremost, Grok isn’t conscious, even with its Companions and Personas. Ultimately, Grok is just a complicated algorithm that takes in prompts and spits out responses, nothing more. So, that means Grok can’t actually be your doctor, friend, partner, or therapist.

Most chatbots explicitly forbid adult content or taboo topics like hate speech, but Grok doesn't. That said, Grok does draw the line at certain subjects, such as anything illegal or the sexualization of children, among others. That said, it can be easy to get around Grok’s restrictions: For example, Grok is happy to point you toward the top torrenting sites if you ask, and it often lets NSFW imagery of real people slip through its moderation cracks.

Outside of policy restrictions, Grok has context window limitations. Think of a context window like a chatbot’s short-term memory, which determines how much information it can process at once. Grok 4 has a context window of 256,000 tokens, and Grok 3 has a context window of 131,072 tokens. This isn’t quite as high as Gemini with its 1,000,000 token context window, but it’s significantly better than ChatGPT’s 32,000 token context for GPT-5. Anecdotally, I didn't hit any context window limits during my testing.

Grok isn’t clear about what its precise usage limits are, and chatbot usage caps are usually dynamic. SuperGrok promises increased access to Grok 3 and Grok 4, and SuperGrok Heavy unlocks extended access to Grok 4. Grok’s API documentation notes a 480 request-per-minute limit for Grok 4 and a 600 request-per-minute limit for Grok 3, but it’s not totally clear if this is only for API access or also for Grok’s paid plans. I didn't hit any usage caps, but you probably will if you use Grok's free plan extensively.

Is Your Data Safe With Grok?

Reading through Grok’s privacy policy, xAI collects significant amounts of data, including, in menu order: account data, payment data, communication data, user content, feedback data, social media information, technical data, location data, publicly available data, internet search data, and cookie data. User content includes all your conversations with Grok, and by default, Grok uses your data to train its models. However, you can turn that off.

In bold type, the privacy policy says the following: “We ask that you do NOT include personal information in your prompts and inputs into our Service; however, we cannot control what you provide to us.” It’s hard not to read this as a cynical disclaimer hidden away in a document few will see, given not just the above but how Grok has doctor and therapist personas that exist to receive personal information.

So, how well does xAI protect all this data it collects? The short answer is not that well. Just this year, a federal government employee leaked a sensitive xAI security key, and a rogue xAI employee tampered with Grok such that it brought up white genocide in South Africa unprompted. Since 2022, X has been dealing with the accusation that 200 million user records were leaked, a story that continues to develop to this day. Going back further, there was the infamous 2020 X account hijacking scandal, too. Whether it’s negligence on the part of an individual connected to xAI or an outright hack, xAI has a sordid security history. 

Accordingly, I suggest following the advice in xAI’s privacy policy: Don’t include any personal information in your prompts.