Grok’s Latest Features Are Wilder Than I Expected — Here’s What’s Actually Worth Using

Written by Shahzaib Ali

Grok’s Latest Features

I’ll admit I was a Grok skeptic for a long time.

When it first launched as an X-exclusive chatbot with a “rebellious” personality and a lot of Elon Musk hype around it, I wrote it off as more marketing than substance. I’m in AI tools practically daily for work — writing, research, coding help, the whole stack — and Grok didn’t feel like it was competing seriously with the tools I was already using.

Then a developer friend told me to actually try it in the first half of 2026, not the version from eighteen months ago. He specifically said “it’s not the same product anymore.” That phrasing stuck with me, because people don’t usually say that unless something genuinely changed.

He was right. It’s not the same product anymore.

Here’s what I’ve actually found after spending several weeks seriously using Grok’s latest features — what’s genuinely impressive, what’s overhyped, and a few things I wish someone had told me before I started.

A Quick Note on What Grok Actually Is Now

One thing worth establishing upfront: Grok has changed structurally in ways that go beyond feature updates. xAI was formally absorbed into SpaceX as the SpaceXAI division on May 6, 2026, with Grok, the X social platform, and the Colossus supercomputer all operating under SpaceX. Nearly all of xAI’s original co-founders departed during that transition. ClickUp

Whether that matters to you as a user depends on your priorities. What it means practically is that Grok now has access to SpaceX’s infrastructure ambitions in a way that shapes its roadmap differently from competitors.

The product itself is available at grok.com as a standalone web app, on iOS and Android, and increasingly embedded in third-party tools. The free tier exists with limited message counts. SuperGrok costs $30 per month with higher limits and access to premium features, and SuperGrok Heavy costs $300 per month targeting enterprise and power users with the highest rate limits. ClickUp

Most of what I’ll describe here is accessible on the SuperGrok tier or below.

DeepSearch: The Feature I Actually Use Every Day

Grok’s DeepSearch is the feature I come back to most consistently, and it does something slightly different from how most people think about AI research tools.

When you enable DeepSearch, Grok doesn’t just answer from training data. It runs a multi-step research process — searching, pulling sources, cross-referencing, and then synthesizing an answer that shows you its work. The process is visible, which matters for research credibility.

I used it extensively for a competitive analysis project. Instead of getting a confident-sounding paragraph I couldn’t verify, DeepSearch returned a structured breakdown with clear sourcing for each claim. It felt more like working with a research assistant than a chatbot.

The limitation I hit: the 10-step DeepSearch limit caps the number of sources consulted per multi-step research task. For simpler research tasks, 10 steps is more than enough. For deep technical deep-dives, I sometimes wanted it to keep going and hit the ceiling at an awkward moment. Suprmind

There’s also an important reliability caveat that I want to be straight about. Grok-3 scored 94% citation hallucination on the Columbia Journalism Review test — the worst of any model tested, with citations generated but the claimed information not matching source content in the majority of cases. This was a real problem in earlier versions. In my recent experience, the situation has improved — but I still treat Grok citations as starting points for verification rather than confirmed facts. Don’t cite something from Grok’s output in anything important without checking the original source yourself. Suprmind

Grok Imagine and Video Generation: The Numbers Are Genuinely Staggering

I wasn’t expecting the video generation to be as far along as it is.

Grok Imagine went from initial release with six-second text-to-video clips in July 2025, through version 1.0 in February 2026 with 10-second clips at 720p, and then gained “Extend from Frame” in March 2026 — the ability to chain clips together using the final frame of one as the start of the next, enabling longer sequences. BASENOR

The “Extend from Frame” feature is what changed video generation from a novelty into something approaching a real workflow tool. Before it, you got standalone clips. Now you can build continuous visual sequences — one clip flowing into the next — which opens up product demos, short narrative videos, and social content that wasn’t viable before.

The platform was generating 1.245 billion videos in January 2026 alone, with 314 million visits to the Imagine feature by early March. These aren’t hobbyist numbers. BASENOR

The latest image-to-video model, grok-imagine-video-1.5-preview, became available via the xAI API in preview as of June 9, 2026. xAI

In my testing, the text-to-video quality is genuinely impressive for short clips with clear prompts. Where it falls apart: complex scene compositions, anything requiring consistent character appearance across multiple shots, and — as noted — quality does degrade when you chain many extensions together. It’s a creative tool for iteration, not a finished production system. Manage expectations accordingly.

Grok Build: The Coding Agent That Came Out of Nowhere

This one genuinely surprised me, because I hadn’t seen it coming from Grok specifically.

Grok Build launched in early beta for all SuperGrok and X Premium Plus subscribers in late May 2026, described as a new coding agent that runs right from your terminal. xAI

The concept: you describe what you want to build in natural language, and Grok Build writes the code, runs it, debugs it, and produces something deployable. It’s in the same category as Claude Code and other agentic coding tools, but with Grok’s real-time data access baked in.

As of early July 2026, the changelog shows active daily development — version 0.2.94 introduced a /goal command, improved session management with an Agent Dashboard, and various fixes around file reading and terminal compatibility. xAI

My experience: it’s fast and handles well-defined tasks cleanly. For “build me a simple API endpoint that does X” type requests, it produces working code quickly. For more complex multi-file projects, it requires more hand-holding than I’d like in its current state. It’s beta software and behaves like beta software — capable but inconsistent.

Grok Build 0.1 became available via the xAI API in public beta on June 1, 2026, so if you want to build applications on top of it, the API access is there. xAI

To get started with Grok Build:

1. Make sure you have a SuperGrok or X Premium+ subscription
2. Install the CLI: the install command is available at x.ai/build
3. Start with small, clearly defined tasks before attempting complex projects
4. Use the /goal command to set a clear objective for each session — it genuinely improves output quality

Memory and Projects: The Features That Make Grok Sticky

These two are less flashy than video generation but more valuable to me in daily use.

Grok Memory is a consumer app feature launched in beta in April 2025 that stores context outside the context window and persists across sessions, with users able to review, edit, and delete memory entries through settings. Suprmind

In practice, this means Grok remembers that you prefer technical explanations without excessive hand-holding, or that you’re working on a specific project across multiple sessions, or that you want responses in a certain format. The memory isn’t perfect — it occasionally surfaces context that isn’t relevant — but having an AI assistant that doesn’t make you re-explain yourself every conversation is a different quality of experience.

Projects work similarly to how Projects function in other AI tools: you create a container with specific context, documents, and instructions, and every conversation within that project operates with that shared context. I’ve found this most useful for ongoing client work where I want Grok to always have relevant background without pasting it repeatedly.

The Voice Features Have Gotten Serious

This was the area I expected least from Grok and found the most surprising.

The Text-to-Speech API became generally available, and a Voice Agent API was launched as well. Voice cloning was added — you can clone a voice from a short audio clip and use it across both the Text-to-Speech and Voice Agent APIs. xAI Docs

The Text-to-Speech API launched in general availability in early May 2026, expanding the developer voice surface beyond the existing Realtime voice and speech-to-text endpoints. Suprmind

For consumer use, Grok’s voice chat has two options — Ara and Grok — and both sound considerably more natural than I expected. For developers, the voice cloning capability opens up applications in content creation, accessibility tools, and customer-facing products.

Frontier voice quality is now also available to Vapi agents through xAI’s infrastructure, which suggests the voice stack is being positioned as a serious competitor in the developer voice API space. xAI

The Microsoft Office Integration Nobody’s Talking About

This came out of nowhere for me and deserves more attention than it’s gotten.

Grok now works inside Microsoft PowerPoint — turning outlines into slides, expanding decks, and tightening narratives without leaving the app — and there’s also a Grok add-in for Microsoft Word to turn notes into documents, style and format work, or bring research from the web into Word. xAI

I tested the Word integration on a client report draft. You highlight a section, ask Grok to improve it or expand it, and it works within Word’s native interface. The web research integration — being able to pull current information directly into a Word document without switching tabs — is the genuinely useful part. It’s available in the Microsoft Store for both apps.

This positions Grok as the AI assistant that lives where a huge portion of professional work actually happens, rather than requiring you to copy-paste between a chat window and your document.

The Companions Feature: What You Need to Know

I’d be doing you a disservice by skipping this because it’s been part of Grok’s public conversation and you may encounter it.

Companions are 3D animated AI characters launched in July 2025 that include Ani (anime-style female companion), Rudy (a friendly red panda), Bad Rudy (a vulgar variant), and Valentine (a male companion). They use the Grok 4 underlying model with real-time image generation and persistent memory across conversations. Some Companions support NSFW mode. Companions require a SuperGrok subscription at minimum. Suprmind

The Companions feature received regulatory and public criticism for its explicit capabilities, with multiple outlets covering the issue and AI Forensics citing the feature in its DSA analysis. Suprmind

For professional use, the standard Grok chat interface with Memory and Projects provides equivalent conversation continuity without the Companions framing. I mention this not to editorialize but because if you’re evaluating Grok for a workplace context, it’s a factor worth knowing about.

Mistakes I Made Getting Started

The biggest one: trying to use Grok for citation-dependent research without verifying sources independently. The early hallucination rate on citations was real, and even with improvements, I’ve caught a citation that pointed to a real publication but paraphrased its findings incorrectly. Always verify.

Second: ignoring the model selection. Grok currently offers multiple model tiers — different versions optimized for speed versus quality. Not selecting the right model for the task is like always driving in first gear. For quick questions, the faster models are excellent. For deep research and complex writing, the more capable models produce noticeably better output.

Third: not setting up Memory preferences intentionally. Memory works much better if you actively tell Grok what to remember in your early sessions — your profession, your preferred response style, what projects you’re working on. Left to its own devices, Memory picks up on things eventually, but guiding it upfront makes the first week much more useful.

Where Grok Actually Stands Right Now

Grok spent the first half of 2025 as a novelty: the chatbot that lived inside X and said things other chatbots would not. What replaced it is a full multimodal platform with standalone apps, enterprise APIs, voice agents, image generation, and video capabilities that arrived faster than anyone outside xAI expected. ClickUp

That’s an accurate summary. What I’d add from using it: the integration breadth is now legitimately competitive — Office plugins, terminal coding agents, voice APIs, video generation, real-time X data access. No single one of these beats the best specialist tool in that category. But the combination, accessible through one subscription, with persistent memory across sessions, is a meaningfully different value proposition than it was a year ago.

The real-time X data access remains genuinely differentiated. For anything where social media sentiment, trending topics, or real-time public conversation is relevant — market analysis, content research, breaking news context — no other major AI assistant can touch what Grok can do with that data access.

The citation reliability is the thing I’d watch most carefully. If that improves consistently over the next few months, the research use case becomes much more compelling. For now, treat Grok as an exceptional research starting point and apply your own verification to anything that matters.

My developer friend was right. It’s not the same product anymore.