Under a YouTube video about Grok Bot, xAI's agent product, a viewer asked: "How do we know GrokBot is using 4.7? Was there a confirmation anywhere?"
Part of the answer you can check in a few minutes. Part of it isn't published anywhere. The useful skill is knowing which is which, and the same steps work whether you're checking a chatbot, a coding assistant, or an agent that works on its own.
A product and a model are different things
An AI model is the part that thinks: a system trained by an AI company to understand language, reason, and write. An AI product is everything wrapped around a model: the app, the interface, the tools it can use, the rules it follows. ChatGPT is a product. GPT-5 is a model. Grok Bot is a product. Grok 4.7 is a model.
Products can use more than one model, and they can swap models without changing their name. That's why questions like this one come up: the product said one thing, and "which model is in there right now?" is a separate fact that has to be checked on its own.
So before checking any model claim, separate two questions: what model can this product use, and what model handled this request? The steps below cover both.
Start with the company's own announcement
The best starting point is also the least exciting: the announcement pages and documentation from the company that makes the product. When xAI released Grok 4.7 in September 2026, the announcement stated that the new model was trained to understand the Grok Bot harness. That's a confirmation, straight from the source, that Grok Bot and Grok 4.7 work together. The viewer's second question, "was there a confirmation anywhere?", had an answer all along.
Notice what that sentence doesn't say: it doesn't say Grok 4.7 answers every prompt inside Grok Bot. Company announcements confirm what the company did. They rarely spell out how a product routes work behind the scenes. Use the announcement to answer "can the product use this model?" and treat "does it use it for everything?" as an open question.
Most AI companies publish model announcements, product changelogs, and release notes. Reading them takes minutes. It won't settle every question, but it settles the one the viewer asked.
Check the model picker
Many products name their models in the interface: a dropdown in the settings, a label next to the answer, a plan page listing what each tier includes. If a tool shows a model name there, you have a useful piece of evidence from the vendor. It tells you what the product says it selected, but it may not reveal how every request is handled.
Watch for variants (different versions or modes of the same model). When xAI released Grok 4.7, it also served a faster version of the same model, available only inside Cursor and Grok Build (its coding products) and not through the public API, the connection other software uses to call the model directly. Same brain, different packaging, different price. The label "Grok 4.7" can be true and still hide which version you got.
Why asking the tool is the weakest check
You can ask an AI which model it is. The answer just doesn't mean much. A model can't inspect its own wiring; it only knows what its instructions say. A November 2024 study of 27 language models found that about one in four misrepresented what they are, and traced the confusion to hallucination rather than deliberate deception. If a tool's instructions say "You are GrokBot running Grok 4.7," the model will say so, whether or not it's true.
Asking can still be worth a try, mostly for what it reveals: if the answer disagrees with the product's own settings page, that mismatch is worth noticing.
Run a small test
Two light tests give you better evidence than any single answer.
Test the knowledge edge. Each model has a training cutoff, the most recent date it learned from. Ask about an event from right before the new model's release, something like "what happened in the news on [that date]?" A model that has never heard of it points to an older version. Two things blur the result: a product with web search or its own reference material can fill the gap, so run this test with search off if the tool allows it, and older models can still guess or hedge their way to a vague answer.
Test the headline feature. New model releases lead with a specific skill: longer tasks, better documents, a new format. Give it one small job in that area. If the launch demo showed multi-hour coding runs and your three-step request falls apart, something is off. Behavioral tests tell you what the product did on that task. They don't prove which model handled it, but they settle the question that matters more: can this tool do your work?
Is that for every prompt?
Agent products often route different steps of a job to different models: reading your request, planning, writing code, checking the result. Some steps are cheap and go to a small, fast model. Some are hard and go to a larger one. This is called routing, and it's one way products keep costs down.
So when a product says it "uses" a top model, the accurate translation is: it can use that model for at least some of its work. The viewer's follow-up question, "is that for every prompt?", usually has no public answer. Vendors often don't publish routing details, and they may change them without changing the product's name. It's a bit like a restaurant that doesn't publish its supplier list: you can know the menu and not the kitchen.
When the label matters, and when it doesn't
Treat the model name as a hypothesis, not a promise. It tells you what the tool is capable of at its best. Your own tasks tell you what it does on your work.
This is also why model comparisons online go sideways. Two products can claim the same model and behave differently, because the harness around the model (its tools, its instructions, its file access) shapes the result as much as the model does. Comments under comparison videos argue about this constantly: did each model get the same tools, the same retries, the same budget? Without those controls, the comparison measures the products, not the models.
The practical version: before paying for any AI tool because of a model claim, run your own task on it. Many products have a free tier or trial. A useful shorthand to keep in mind: the model sets a rough ceiling on what's possible, and the product around it decides how much of that ceiling you reach.
Frequently Asked Questions
How do I know which model ChatGPT is using right now?
OpenAI names the model in the interface and documents model changes in its release notes. The picker in ChatGPT shows which model is selected for the conversation, and the product may adjust speed and thoroughness with different modes. For anything that matters, OpenAI's own announcements are the source to check.
Can an AI model lie about which model it is?
Deception on purpose is rare; misstatement is common. Models answer identity questions from their instructions and training, not from any ability to check what they're running on. The November 2024 study mentioned above found roughly one in four models gave a wrong or inconsistent self-description, and traced it to hallucination rather than intent. Treat self-reports as a hint, then verify with the vendor's documentation.
Why did my AI tool get worse overnight?
A few likely reasons: the product switched which model handles your task, changed the settings around it, or moved some tasks to a cheaper model to manage cost. The product name and interface stay the same while the model behind them changes. Check the vendor's changelog first; if other users noticed the same shift, the dates there tend to line up.
What is model routing?
Model routing is when a product automatically picks which AI model handles each part of a task, trading cost and speed against capability. It's why a tool can claim a top model and still not use it on every request.
Want to go further?
- How to Verify AI Output: A Simple Framework covers the habit of checking what AI tells you, model labels included.
- What Is an AI Agent? Explained Simply covers how agent products work under the hood.
- ChatGPT Free vs Paid: Is the Upgrade Worth It? covers when to pay for a tool and when the free tier is enough.