What is Grok 3, and what do its Think and DeepSearch modes do?

Last night, Monday, February 17, 2025, Elon Musk and members of his AI company xAI presented Grok 3 in a livestream on X, the social network Musk owns. Grok is xAI's chatbot and the AI behind several features on X. The new version arrives a little later than first hoped: it had been expected in 2024.

Musk's pitch was bold. He called Grok 3 "an order of magnitude more capable" than Grok 2, meaning roughly ten times better, and described it as a "maximally truth-seeking AI". He also said it is still in beta and will improve almost daily. Here is what was actually announced, and how to read the claims.

What xAI announced

Grok 3 is not a single model but a family, rolling out from Monday, with some parts still in testing:

Model or feature	What it does
Grok 3	The main model, answering questions and analyzing images
Grok 3 mini	A smaller version that answers faster, with some loss of accuracy
Grok 3 Reasoning, plus a mini Reasoning version	Models that "think through" problems before answering
Think	A button in the Grok app that switches on step-by-step reasoning
Big Brain	A mode that spends extra computing on the hardest questions
DeepSearch	Searches the web and X, then delivers a summary answer

The upgrades appear in Grok's iPhone and web apps. Musk said a voice mode should follow in about a week, and that developers will get Grok 3 through xAI's API a few weeks after that. He also said xAI plans to release Grok 2 as open source once Grok 3 is stable, likely within a few months.

How it was built

The headline is scale. TechCrunch reports that xAI trained Grok 3 in a huge data center in Memphis holding about 200,000 GPUs, the graphics chips used to train AI. Musk says the model used around ten times the computing power of Grok 2, and a larger training set that includes court filings, among other material.

The reasoning models work like OpenAI's o3-mini and DeepSeek's R1. A reasoning model does not reply immediately. It first writes out a hidden chain of steps, checks its work and revises, then gives the answer. That makes it slower but more reliable on math, science and programming.

Think of the difference between answering a math question out loud on the spot and working it on scratch paper first. Think gives Grok the scratch paper; Big Brain gives it more paper and more time. Musk said some of these intermediate "thoughts" are hidden in the app to prevent distillation, a technique in which one company trains its own model by copying another model's outputs.

DeepSearch combines reasoning with search. It reads web pages and posts on X, then writes a summary. Musk described it as doing in about ten minutes the kind of research that might take a person an hour. It is xAI's answer to research tools such as OpenAI's deep research.

What the benchmarks show, and what they don't

During the stream, xAI showed charts in which Grok 3 beats GPT-4o and DeepSeek-V3 on tests of math, science and coding, including AIME, a competition-math exam, and GPQA, a set of graduate-level science questions. xAI also says its reasoning model beats o3-mini at its highest setting on several tests, including AIME 2025, a new set of math problems, and that an early version of Grok 3 did well in Chatbot Arena, where people vote on which of two anonymous answers they prefer.

Two cautions apply. These are xAI's own numbers, chosen and presented by xAI. And it is very early: reactions on X range from impressed to skeptical. Andrej Karpathy, a well-known AI researcher, wrote that the benchmark results look encouraging, which is a useful signal but not a full independent evaluation.

Try it: Pick three questions with answers you can verify: a math word problem, a short coding task and a question about this week's news. Ask each one to Grok 3 with Think on, and to another assistant you use. Write down which answers were right, which explained their steps, and how long each took.

Why it matters

Grok 3 shows how crowded the top of the field has become in just a few weeks. DeepSeek's R1, OpenAI's o3-mini and OpenAI's deep research have all arrived recently, and now xAI joins with a model trained on a cluster of about 200,000 GPUs. The common thread is reasoning and research: every major lab now offers some version of "think longer" and "go search for me".

It also continues Grok's particular identity. Since its first release in November 2023, Grok has been promoted as willing to take on questions other assistants refuse, and Musk keeps presenting it as an alternative to chatbots he criticizes as too "woke". Whether Grok 3 is more balanced, as Musk has pledged, is not yet clear.

Who can use Grok 3 right now, and how should you test it?

Grok 3 goes first to subscribers of X Premium+, priced at $50 a month according to TechCrunch. xAI also announced a separate plan called SuperGrok for the Grok app and website, with more reasoning and DeepSearch requests and unlimited image generation; its price is reported, but not confirmed, at $30 a month or $300 a year. Some features are in beta and may change from day to day.

If you have access, test it the way you would test any new assistant. Use Think for problems with one right answer, and check the result yourself. Use DeepSearch for a question where you already know the main facts, then open its sources to see whether they really support what it says, especially when the sources are posts on X rather than established publications. Keep personal and sensitive information out of your prompts.