Grok's agents are no longer a $300-a-month thing: what they are and what they do when you look away

For three years, talking to an artificial intelligence has meant the same thing every time: you ask, it answers, and the conversation stops there until you type again. What xAI has been quietly shipping over the past two weeks breaks exactly that part, because an agent does not wait for your next question, it keeps working after you close the laptop and go and take a shower.

Grok 4.6 landed on August 12, and we already covered its benchmarks and its price. But the interesting part arrived afterwards, buried in release notes: on August 19 Grok Build opened to every plan on web and mobile, and on the 21st access to Grok Bot stopped being locked behind the $300-a-month subscription.

What an agent is

A normal language model is a very clever person locked in a room with no doors: it knows an enormous amount, reasons well and will write you anything you ask for, but it cannot touch the outside world. Ask it to book a flight and it drafts you a lovely email for you to send yourself.

An agent is that same person handed the keys to a computer.

Same brain, mind you, but now with a browser where you are already logged in, a terminal to run commands in, and permission to chain steps together for hours without checking in every two minutes. The practical difference is not how clever it is, it is what it can do while you are not there, and that changes what you can reasonably ask for, instead of "write me the email", it becomes "go through this month's invoices, flag the ones that do not add up and leave me the list by morning".

That is what xAI is selling with Grok Bot, you talk to it in a chat, the way you would to a colleague, you say what you want and it organises itself. Each bot runs on its own machine in the cloud, can log into tools and websites even when they have no API, and several bots hand work to each other, escalating to you only what genuinely needs a human decision.

What it does once your back is turned

Grok 4.6 was trained with this in mind. Its biggest gains sit in the long-horizon work benchmarks rather than the knowledge ones: on DeepSWE v1.1 it climbs from Grok 4.5's 54% to 65.9%, and on Terminal-Bench v3.0 it jumps from 15.7% to 26%, as reported by TestingCatalog. Both tests measure whether a model can hold a many-step task together without derailing, which is precisely what an agent needs to avoid leaving you stranded halfway.

"It delivers frontier intelligence and is a significant improvement over Grok 4.5 at the same price," SpaceXAI writes in its announcement, and the price part checks out, the API still costs $2 per million input tokens and $6 per million output tokens, exactly what the previous version cost. On the Artificial Analysis index the model rises from 56 to 61 points, tying GPT-5.6 Sol Max.

Grok Build is the more tangible half for anyone who does not code. You describe an app, it builds it, and since August 19 you can publish it to a grok.me link, point your own domain at it or export it to GitHub without paying extra. If you want to try any of this on a Spanish budget this afternoon, start there, it is the only piece that does not ask for a $300 plan.

Where the small print lives

This is where it pays to slow down. Grok Bot is still in beta, and beta in 2026 means it does odd things and the bill moves. Subscriptions bundle a weekly allowance and anything beyond it bills separately, by model and tokens burned. An agent grinding away for eight hours eats far more tokens than you ever will asking it questions.

The second warning is more serious. An agent that walks into your tools with the session already open is an agent that can be fooled, that is, if it reads a malicious web page halfway through a task, that page can try to give it instructions. The industry has been wrestling with this for months, though it says little in public, and we covered the alliance Nvidia assembled to attack exactly that problem.

At INSERT FUTURE we think that within 6 months "agent" will stop being a feature anyone announces and will simply be how every assistant works, the way nobody advertises that their phone has a camera any more. I hope I am wrong about the timing, because the security side is still lagging behind.

An intern who never sleeps is astonishingly cheap. Right up until it breaks something and bills you by the hour.