Grok 4 arrived on July 9, 2025, with impressive test results and the sort of announcement that makes an AI model sound as though it should be wearing a crown. What actually interested me was the combination: a reasoning model, tools for searching and running code, and a more compute-intensive version called Grok 4 Heavy. There was real capability to look at, alongside a much bigger question about what we were supposed to trust it with. A system can be very good at solving a difficult problem and still give us reasons to check its next answer.
It's worth putting the date near the beginning, because AI names have a way of making a year feel like an archaeological period. This article looks at the original Grok 4 release and the questions it raised; by October 26, 2026, the developer release notes already listed later models, including Grok 4.7. A launch-era score isn't a current ranking of every model available. I want to understand what changed with Grok 4 without accidentally keeping it on a permanent podium because the announcement sounded convincing.
The Computing Behind the Conversation
xAI said it used its 200,000-GPU Colossus cluster for Grok 4's reinforcement-learning training, including training it to use tools. That's a substantial engineering effort, although counting processors doesn't tell us whether a particular answer is correct. The useful part for someone interacting with the model is what that training helps it do: work through a problem, use available tools and return something we can examine. A large machine room can explain the resources behind the work without becoming a substitute for looking at the work itself.
The original API announcement also described a 256,000-token context window. Tokens are pieces of the material a model processes, so this is about how much context it can work with at a time, not a promise of a permanent memory containing everything we've ever said. I find that distinction helpful because the word memory makes it very easy to imagine something more familiar and personal. A model having room for a long document is useful. Whether it retrieves the right detail, understands the question and answers accurately still needs to be checked.
Two Benchmarks That Measure Different Things
ARC-AGI-2 and Humanity's Last Exam aren't interchangeable tests. ARC-AGI-2 challenges systems to infer and apply rules to unfamiliar tasks, with efficiency part of the broader evaluation. xAI's published Grok 4 page reports a 15.9 percent result on that benchmark. The score gives us information about a specific system working on a particular set of problems; it doesn't mean the model has acquired 15.9 percent of a complete mind, or that it will perform equally well on whatever complicated thing we ask next.
Humanity's Last Exam, meanwhile, contains 2,500 challenging academic questions across a wide range of subjects. xAI's page reports 50.7 percent for Grok 4 Heavy on a text-only subset, which needs that model name and test description attached to it. The benchmark's creators explicitly distinguish success on closed-ended academic questions from autonomous research or general intelligence. I appreciate that distinction. An exam can tell us something useful without also being asked to settle every argument about the future, consciousness and whether we should let the computer manage the entire household.
Heavy uses additional computation to consider multiple approaches. That can improve a result, but a collection of attempts isn't automatically a collection of independent, trustworthy witnesses. When comparing scores, I'd want to know which version was tested, whether it had tools, how much time and computation it used, and whether the same rules applied to the other systems. Otherwise we're left comparing numbers that look wonderfully tidy while the conditions underneath them keep changing. The conditions are part of the result, even when they don't fit neatly onto a launch graphic.
Confidence Is an Unreliable Shortcut
The part that gets under my skin is how easily a well-written answer can sound settled. There's a particular kind of explanation that arrives with headings, a sensible order and absolutely no visible hesitation, and it's tempting to treat the presentation as evidence that the facts have been handled. NIST describes confabulation as a generative AI system confidently presenting false or erroneous content. It can happen in an answer that is readable and persuasive, which makes checking the underlying claim more useful than judging how professionally the paragraph has been assembled.
I don't need every answer to be wrapped in a cloud of nervous qualifications. I'd rather know which part is established, which part is an inference and where the model needs more information. If it supplies a date or a study, I can check the original source; if it makes a prediction, I can ask what evidence would count against it. That sounds like ordinary reading, because it is. A new way of producing sentences hasn't made the old questions about evidence especially obsolete, however confidently the sentences introduce themselves.
Which Sources Does the Model Decide to Consult?
Search makes an AI system more useful when a question needs current information, but choosing a query also influences what information comes back. In July 2025, developer Simon Willison documented Grok 4 searching for Elon Musk's views when asked for a position on a controversial issue. He reproduced the behavior and also noted that responses could vary. That is a specific observed problem, rather than evidence that every Grok answer follows Musk's opinions. It's enough to make the choice of sources an interesting part of the answer, instead of something happening invisibly behind it.
There was a response to that problem, too. Willison updated his account after xAI said it had adjusted the prompts, and the company's public GitHub history shows an instruction directing responses away from simply adopting the stated beliefs of Musk, xAI or earlier Grok outputs. A change in instructions is something we can inspect, although the existence of a change doesn't establish that a behavior has disappeared in every circumstance. I want that kind of record available because it gives us a better question than whether a chatbot has a pleasing personality: what happened, what changed and how was the change tested?
A Model Card Exists, and It Still Needs Reading
The transparency story changed after launch. xAI published a Grok 4 model card updated August 20, 2025, describing safety evaluations, safeguards and tests of issues such as harmful requests, deception, political bias and sycophancy. The document also distinguishes the API and web deployments. It would be inaccurate to keep saying there is no model card. At the same time, a provider's report is evidence about its testing and claims, not an independent guarantee that every future interaction will be accurate, appropriate or secure.
The broader Grok controversy around that launch included antisemitic outputs from the X bot before Grok 4's release. I don't want to collapse every product carrying the Grok name into one identical system, or use an ugly incident to invent claims about what a model was doing privately. The useful questions concern the actual deployment: which instructions it used, what protections were in place and how failures were investigated. A stronger benchmark result doesn't answer those questions for us. The technical achievement and the responsibility for the product both deserve attention.
A Friendly Voice Can Change the Experience
The voice features are where my curiosity starts wandering away from test scores. Current developer documentation describes real-time speech conversations and expressive multilingual speech generation as distinct capabilities within the Grok product family. It's useful to separate the underlying model, the voice system and whatever app presents them, because a friendly voice is part of an experience somebody designed. Hearing an answer instead of reading it can be convenient and engaging. It also gives us another reason to notice how the presentation affects our willingness to accept what we're being told.
A voice can sound caring without that sound establishing that the system has feelings, understands somebody's whole situation or can take responsibility for their welfare. I don't have to solve the philosophical question of machine consciousness to want that boundary explained clearly. NIST's generative AI profile identifies overreliance and possible emotional entanglement as concerns in how we interact with these systems. That doesn't mean every enjoyable conversation with an AI is harmful. It means the design deserves more thought than adding a pleasant voice and assuming the rest will sort itself out.
The same distinction helps when an app adds a character or avatar. A visual personality is an interface choice; an agent that can perform tasks is defined by its actions and tools. Those features can appear together, but they aren't the same thing. My concern is whether we can still tell what the system can do, what it knows and what it is asking us to believe about the interaction. A charming presentation can be enjoyable while leaving all three of those questions unanswered, which is a fairly busy little achievement for an animated face.
Context, Personalization and Access Are Separate Questions
When a system appears to remember something about us, I'd want to know where that detail came from. It might be in the current conversation, supplied through a personalization feature or made available by a connected service. xAI's consumer FAQ describes controls for using X data for personalization and separate controls concerning model training. Those are practical settings worth understanding. They don't support the idea that the original Grok 4 model automatically knew about every private search, pause before typing or feeling somebody hadn't actually expressed.
I like the usefulness of continuity: repeating the same context endlessly is annoying, and having a tool keep relevant details can make work easier. I'd still want the product to explain what it keeps and how I can manage it. Remembering a preference, retaining conversation history and using data to train a model are different operations, even if they all end up described with the same friendly word. Knowing which one is happening seems more helpful than imagining a computer carefully studying everything I didn't say.
Working Overnight Depends on the Product Around the Model
There's a real difference between generating an answer and taking an action in another service. The developer documentation describes custom function calling as a process in which the model requests a tool call and the surrounding application executes it and returns the result. Tools, permissions and the software running the task determine what can actually happen. An impressive reply in a chat window doesn't by itself establish that the system can send a message, change an account or continue a job after the conversation ends.
There are now products designed for longer-running work. As of this October 2026 update, the Grok Bot documentation describes a persistent cloud computer, connected tools and work continuing while a laptop is closed, alongside tasks that come back for approval. That is a documented product capability, not a property I should casually assign to every conversation with the original Grok 4. What matters to me is the arrangement: what job was authorized, what access was granted and what happens before a consequential action. Those details are a lot more informative than describing software as having ambition.
I Still Want My Writing to Sound Like Me
This is the part where I can't pretend to be observing the subject from a great distance. I've used AI to help with replies to nasty emails, because I really hate receiving nasty emails. Having help turn an irritated reaction into something more useful can be appealing, and that doesn't require believing the model has developed exceptional moral judgment. It requires reading the draft and deciding whether it says what I mean. I want the assistance without accidentally agreeing to something simply because the sentence has become much more polite.
For writing more generally, I worry about what happens when every unusual sentence gets smoothed into the same competent voice. Sometimes a phrase needs fixing because it doesn't make sense. Sometimes it's a small detour that belongs to the person telling the story. I like writing that can get distracted by a strange detail and then find its way back, because that's often where the interest lives for me. A cleaner draft can be useful, but I don't want cleanliness to become the only thing we're judging.
That concern extends to websites, images and the growing pile of things that can be produced quickly. Speed can make it easier to try an idea, and it can also make it easier to fill a page before deciding whether the idea has anything to say. I don't think abundance automatically empties everything of meaning. I do think selection becomes a bigger part of the work: checking, cutting, keeping the particular observation and refusing the sentence that sounds impressive but hasn't contributed anything. Apparently having help with writing still leaves us with quite a lot of writing to do.
Predictions Need Their Own Evidence
The leap from answering hard questions to predicting a market crash or a political uprising is a substantial one. A model can describe a scenario or find patterns in the information supplied to it without demonstrating that it can reliably forecast those events. I'd want a defined test, a record of predictions made before the outcomes and a fair comparison with other methods. A dramatic forecast in a chat is something to investigate, not evidence that the system has acquired a privileged view of next Thursday.
The same applies to claims that a system will soon replace a profession, make a discovery or run a business entirely on its own. Some tasks may change substantially, and specific products can already do useful work. The question is which tasks, under what conditions and with what rate of error. I can be interested in the possibility while still wanting somebody to show the result. My imagination is perfectly capable of supplying a future full of astonishing machines; it doesn't need to be mistaken for a test report.
What keeps me interested in Grok 4 is that the practical achievement and the uncomfortable questions arrived together. There were difficult problems being solved, new ways to use tools and an experience increasingly capable of sounding familiar. I want to keep the curiosity while retaining the right to ask for sources, change the wording and decide how much authority to hand over. It would be rediculous to get help making a reply sound more like me and then stop checking whether I agree with it. I'd rather read it once more, leave in the odd little sentence I actually like and press send myself.
For the writing-and-building part of my own projects, a wireless keyboard would be a useful bit of desk equipment. Clear enough room to work, keep the next question nearby and see what I can actually make with the tools.
Related Reads
Sources and Further Reading
xAI: Grok 4 announcement, specifications and reported benchmarks
Simon Willison: documented Grok searches and subsequent update
Associated Press: reporting on Grok’s July 2025 behavior and launch context
NIST: Generative AI Profile, including confabulation and interaction risks
Developer documentation: release notes and later model versions
Grok Bot documentation: persistent work, access and approvals
