AI ·
Grok
The assistant from xAI. Where it runs, how the model line was named, and what Grok 4.7 is as of October 2026.
Grok is the assistant. xAI announced it on 3 November 2023 and described it as an AI modeled after the Hitchhiker's Guide to the Galaxy: built to answer almost anything, and, harder, to suggest the question you should have asked. The word is older than the product. Robert Heinlein used "grok" in Stranger in a Strange Land for an understanding that goes all the way through. The product took the word, and it took the guide's manners.
xAI was founded by Elon Musk in 2023. On 2 February 2026 SpaceX acquired xAI. News pages after that date say SpaceXSI. The API, the documentation, and the install scripts still live at x.ai. I use the name on the door I am knocking on.
Where it runs
On the web at grok.com, in the Grok apps for iPhone and Android, and inside X. On X the account @grok will answer in a thread when you ask it. That public bot is the chatbot sitting where the people are. It is not Grok Bot, the later product, which is a teammate with a computer of its own. The names rhyme. The objects do not.
A free allowance exists. Paid tiers, including X's subscription and SuperGrok, raise the limits and open the heavier tools. Prices move. I do not print a number on a page that will still be here next year. Look at the current plan page when you are about to pay.
What you can ask it to do
Conversation, search over the web and over X, writing, code, images, and voice. The live search is the part that feels different in the chat. Ask about something that happened this morning and the product can go and look, on the web and on X, instead of answering only from the month the training cut off. That look is a tool call. The model then reads what came back. It can misread it. A retrieved paragraph is evidence, and it is still text.
Images and short video live under Grok Imagine. Code that touches a repository, in the terminal, is Grok Build. The chat is the front door to all of it. The specialized tools are where the work gets a harness: files, diffs, a plan you can reject, a computer that stays on.
The model line
These are the dates on the x.ai news pages.
- 3 November 2023. Grok is announced.
- 17 March 2024. The weights and architecture of Grok-1 are released. It is a mixture-of-experts model with 314 billion parameters.
- 28 March 2024. Grok-1.5, with a context of 128,000 tokens.
- 12 April 2024. Grok-1.5 Vision, the first multimodal preview.
- 13 August 2024. Grok-2 and Grok-2 mini.
- 4 November 2024. The API enters public beta.
- 9 December 2024. Aurora, an image model, inside Grok on X.
- 19 February 2025. Grok 3.
- 9 July 2025. Grok 4, with tool use and live search, plus a Heavy tier.
- 17 November 2025. Grok 4.1, on grok.com, on X, and in the apps.
- 19 November 2025. Grok 4.1 Fast, aimed at agents that call tools.
- 16 July 2026. Grok 4.5, for coding, agent tasks, and knowledge work.
- 12 August 2026. Grok 4.6, with the emphasis on long-running agents.
- 21 September 2026. Grok 4.7.
Grok 4.7 is the one I reach for as of this writing. The launch post calls it their most capable model for coding and knowledge work. Two sentences on that page are easy to mix up, so here they are apart. Against comparable models, they describe 4.7 as twice as fast at half the price. Against their own 4.6, they say it is served at the same price and the same speed. The published API price starts at $2 per million input tokens and $6 per million output tokens. A fast variant costs twice that and produces output tokens twice as quickly.
On their table, CursorBench 4.0, a test of longer coding tasks, is 46.3 percent for 4.7 and 40.4 percent for 4.6. Those are the lab's numbers, on a test the lab is willing to quote. I read them as a press sheet that bothered to be specific. I still run the task I actually have.
4.7 is in Cursor, in Grok Build, and on the API, and through the usual routers and clouds. The company says it was trained to understand the Grok Bot harness. That is a 2026 sentence. The model and the agent around it ship as a pair. A benchmark of the bare model and a week of using the agent are different measurements, and a buyer needs both.
The manners
Grok is written to be direct, to joke, and to answer questions other assistants refuse. A personality is a system prompt plus a pile of preference data. It is a choice about tone. Wit and accuracy are different axes. When I want a fact, I ask where it came from, and I open the page. I do the same with every other model. The rebellious streak is not a citation.
If you are deciding whether to build on it, start in the chat until you know the texture, then move the real work to the API or to Grok Build, where you can see the tools and keep the diff. The shelf continues with Grok Build, Grok Bot, and Grok Imagine.