How to Use Grok 4: Access, Features, and Limits

How do you use Grok 4?

You use Grok 4 the same way you use any current Grok model: sign in at grok.com on the web or open the iOS or Android app, then type a question. There is no purchase step: xAI's free tier includes the current flagship model, real-time web and X search, voice mode, and connectors, and you can be asking your first question inside a minute. Paid tiers lift rate limits and add image and video generation rather than gating the model itself.

One clarification on the version number, since "how to use Grok 4" is still how most people search for this. Grok 4 shipped in mid-2025 and has been succeeded; xAI's flagship as of July 2026 is Grok 4.5, and its documentation recommends that model for both chat and code. You do not select Grok 4 anywhere. You sign in and get the current model, which is the one every instruction below describes.

The three-step version fits in a sentence, but the useful part is what comes after: knowing which mode to use for which job, how to make search actually fire, and where the model stops being reliable. That is what the rest of this guide covers.

1. 1Go to grok.com on the web, or install the Grok app on iOS or Android. You can also use Grok inside X if you already have an account there.
2. 2Sign in with an X account or an email address. Your history syncs across every device you sign in on.
3. 3Type your question in the composer. Grok answers from its training by default and reaches the live web and X when you ask it to search or when the question clearly needs current data.
4. 4Switch modes for the job: search for anything current, deeper reasoning for hard problems, voice for hands-free, Imagine for images and video.
5. 5Upgrade only when you hit limits. Free covers the same model, so paying buys headroom and generation features, not better answers.

How to access Grok: every path

There are four ways in, and they are genuinely different products rather than skins on the same thing. Picking the wrong one is the most common reason people conclude Grok cannot do something it can do.

The standalone app at grok.com is the full experience: modes, connectors, Imagine, file uploads, memory across chats, and a workspace that behaves like a tool rather than a feed feature. Grok inside X is lighter and lives in the context of the timeline, which makes it excellent for asking about a post or a trend and cramped for sustained work. X Premium at $8 a month raises your Grok limits inside X, and Premium+ at $40 raises them further.

The API is for software. It has no interface at all, bills per token rather than per month, and is what you reach for when the thing asking questions is your own product rather than a person. If you are evaluating Grok to power something customer-facing, the API is your path and none of the subscription tiers are relevant.

What you can do on the free tier

More than you would expect. xAI's free plan lists real-time web and X search, voice mode, connectors, and SOC 2 Type I and II compliance, and the plan comparison shows the current flagship model available on every tier including Free. That makes the free tier a real evaluation of the product rather than a trailer for it.

Two things are genuinely gated. Search depth is marked "Limited" on Free against "Expert" on paid tiers, so complex multi-source research runs shallower. And image and video generation through Imagine sits behind the paywall entirely, so if visual output is why you are here, the free tier will not show you what you came for.

The third constraint is rate limits, and it is the one that actually pushes people to pay. Free usage is capped, and heavy days hit the ceiling. The right way to use this is deliberate: work on the free tier for two weeks, count the times you get cut off, and let that number decide the $30 question rather than a feature comparison.

It is worth being explicit about why this matters, because most AI vendors do not structure things this way. When the flagship model sits behind the paywall, a free tier is a demo of a weaker product and tells you nothing about what you would be buying. When the flagship model is on the free tier, as it is here, your two-week trial is the real thing. Take advantage of that rather than upgrading on the assumption that paid answers are better answers.

• Included free: current flagship model, real-time web and X search (limited depth), voice mode, connectors.
• Gated: image and video generation via Imagine, Expert-level search depth, higher rate limits.
• Available on web, iOS, and Android with synced conversation history.
• No credit card required, which makes it a genuine two-week evaluation rather than a trial.
• If answer quality disappoints on Free, upgrading will not help; it is the same model.

Getting around the Grok interface

The interface is a composer, a thread list, and a set of mode controls. You type into the composer, the answer appears in a thread, and follow-ups continue with the full context intact. Threads are the unit of work: keep one topic per thread and you get better answers, because the model is reading everything above your question as context.

Attachments and uploads matter more than people use them for. Grok accepts file and PDF uploads for summarizing and analysis, and accepts images for vision tasks like reading a screenshot, a diagram, or a photo of a document. The single highest-return habit is to stop describing a thing you could just paste in.

Three interface details are worth setting up once. Custom instructions let you fix tone and format globally so you stop repeating them. Memory carries preferences and past conversations forward across threads. Connectors bring your own sources into range. Ten minutes on these three at the start pays back over every subsequent conversation.

• Composer and threads: one topic per thread, follow-ups keep full context.
• File and PDF upload for summaries and analysis; image upload for vision tasks like screenshots and diagrams.
• Custom instructions to set tone and output format once instead of every message.
• Memory across chats, so preferences and prior context carry forward.
• Canvas-style long-form editing for drafting with inline revision.
• Shareable conversation links when you need someone else to see the thread.

Grok's modes: search, reasoning, voice, Imagine

Grok's modes are not cosmetic. Each one changes what the model does before it answers, and matching the mode to the job is the difference between a useful answer and a confidently stale one.

Grok features worth knowing

Beyond the headline modes, a handful of features change how much work you can actually move through Grok. The multi-agent capability is the most distinctive: xAI describes several agents working in parallel on sub-problems, each showing auditable reasoning, with the results merged into a single cited answer. That sits on the higher consumer tier rather than the standard one.

The 500,000-token context window on Grok 4.5 is the quiet workhorse. It means you can put a lot of source material in front of the model at once rather than chunking it, which removes a category of error where the model answers from a fragment it happens to have and misses the part that mattered.

Language coverage and cross-device continuity round it out. Grok handles dozens of languages natively, syncs history across web, iOS, and Android, and supports shareable public links for any thread. None of these are headline features; all of them show up in daily use.

Connectors deserve a mention on their own because they are listed on every tier including Free, which is unusual. They let you point Grok at your own sources for a conversation instead of pasting content in by hand. Useful as that is, understand what it is not: a person choosing to bring a source into one conversation. It is not a standing integration, and nothing a connector does is visible to anyone but the person who set it up.

How to get better answers out of Grok

The gap between a mediocre Grok answer and a good one is usually in the question, not the model. The habits below are unglamorous and they work, and they work on every frontier assistant rather than just this one.

1. 1Say what you want the output to look like. "Give me a five-row table comparing these on price, context window, and rate limits" beats "compare these" every time, and it removes a round trip.
2. 2Paste the source material instead of describing it. With a 500,000-token context window there is rarely a reason to summarize an input for the model when you can hand it the whole thing.
3. 3Force search explicitly when the answer depends on current facts. The model has a training cut-off and will answer from stale knowledge if you do not push it to look.
4. 4Click at least one citation before you use the answer. Live search reduces hallucination; it does not eliminate misreading, and a cited answer you have not checked is still unverified.
5. 5Set custom instructions once for tone, length, and format rather than repeating them in every message. This is the single highest-return five minutes in the product.
6. 6Keep one topic per thread. Mixed threads degrade answers because everything above your question is context the model is reading.
7. 7Ask it to show its reasoning on anything where being wrong is expensive, then read the reasoning rather than just the conclusion.

Grok's limitations and where it gets things wrong

The most important limitation is documented by xAI itself: without server-side search tools enabled, Grok has no knowledge of current events or data beyond its training cut-off. For Grok 4.5 that cut-off is 1 February 2026. An assistant that reads the live web is still an assistant that will confidently answer from memory if you do not make it look.

The second limitation is private information. Grok does not know your order database, your return policy, your inventory, or your customer's account status, and no subscription tier changes that. Connectors help you bring sources into range for a conversation, but that is a per-conversation act by a person, not an integration your customers can use.

The third is the one every general-purpose assistant shares: fluent wrongness on the long tail. Industry accuracy benchmarks consistently find that assistants perform well on broadly documented general knowledge and much worse on obscure facts, where error rates climb into double digits. Grounding an assistant in verified sources is what closes that gap, and it is a configuration decision rather than a model-quality one.

• No knowledge past its training cut-off unless search tools are enabled for the request.
• No access to your private business data: orders, policies, inventory, account records.
• Rate limits on the free tier, which is the usual reason people upgrade rather than answer quality.
• Confident errors on niche and long-tail facts, the standard failure mode for general-purpose assistants.
• Not a customer-facing product: no widget, no ticket routing, no human escalation path.

Using Grok through the API

If the thing asking questions is your software rather than a person, you want the API and none of the subscription tiers apply. xAI's API is usage-based with no minimums, and Grok 4.5 is billed at $2.00 per million input tokens and $6.00 per million output tokens, with a 500,000-token context window. It supports tool calling and agentic workflows, server-side web and X search, file storage with retrieval-augmented generation, and batch processing.

Model aliases are the detail worth getting right up front. xAI documents that the bare model name points at the latest stable version, that a "-latest" suffix tracks the newest release, and that a dated name pins a specific version permanently. Use the alias for features, pin the date when a change in behaviour would break something downstream.

The endpoint is OpenAI-compatible in shape, so a request looks familiar if you have used any modern chat completion API. The code below is the minimum viable call; everything interesting comes from adding tools, search, and your own retrieved context on top of it.

const res = await fetch("https://api.x.ai/v1/chat/completions", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Authorization: `Bearer ${process.env.XAI_API_KEY}`
  },
  body: JSON.stringify({
    model: "grok-4.5",
    messages: [
      { role: "system", content: "Answer only from the provided context." },
      { role: "user", content: "Summarize this policy in three bullets." }
    ]
  })
})

• Usage-based billing with no minimums; separate from any Grok subscription you already pay for.
• Tool calling and agentic workflows, plus server-side web and X search as opt-in tools.
• File storage and retrieval-augmented generation for grounding answers in your own content.
• Batch processing for high-volume jobs that do not need a live response.
• Model aliases: bare name for latest stable, dated name to pin behaviour.

Using Grok for customer support (and where it stops)

Grok is good at the reading-and-writing half of support. Drafting a reply, summarizing a long thread, translating a message, checking tone before you send: those are real time savings for a support team, and a $30 SuperGrok seat pays for itself quickly if your agents use it daily.

It stops at the part that actually reduces ticket volume. A customer asking where order 10482 is needs a live lookup against your store, not a language model's best guess. A customer asking whether a jacket bought six weeks ago is still returnable needs your policy applied to their specific order date, with an answer the business will stand behind. Neither is a model-quality problem, and no tier upgrade solves either, because the missing piece is a connection to your systems and a rule about what the agent is allowed to decide.

This is also why per-seat pricing is the wrong shape for support. A seat licence assumes value scales with the number of employees typing. Support volume scales with customers, which is why a store with two support staff and forty thousand monthly orders gets almost nothing from buying two more AI seats. What varies month to month is conversations, so that is what the price should track.

Bookbag prices on exactly that basis: flat monthly plans with message-credit allowances and a spend cap you set, rather than per-seat licences or per-resolution fees. One credit equals one AI reply on any model, a typical conversation runs about four replies, so plans map to conversation volume: Free covers 50 credits, Starter is $30 a month for 600, and Growth is $110 a month for 5,000. The agent connects to Shopify, WooCommerce, or BigCommerce so order questions become lookups rather than guesses, and it hands off to a human with full context when a question falls outside what the data supports. That combination, not a bigger model, is what turns an assistant into a support agent.