I run my companies on Grok 4.6. Not as a test. It is the model behind the agents that handle my research, my operations, and my client follow up, and on my X Premium+ plan it costs me nothing extra to run. So when that model shows up in a week of AI headlines, I am not reading coverage. I am reading about the thing already sitting in my stack.
Here is what happened and what it actually means if you own a business.
xAI launched Grok Bot on August 11
AI teammates that get their own cloud computer and sign into your existing apps and websites. The pitch on human oversight is that the bot comes back to you when it needs approval, instead of asking you to watch every click. Enterprise access is waitlisted. I have seen pricing chatter on Hacker News but nothing confirmed, so I am not going to quote you a number.
The demo is an employee that logs into your stack. That is what most owners thought they were buying the first time they opened ChatGPT. The gap is that a raw model with a browser is not an employee. An employee has a job description, a process, someone checking the work, and a way to escalate when something breaks. Building that layer is the actual work. It is also the part we install.
Grok 4.6 landed the next day
It is an increment on 4.5, aimed at longer running agent workloads, and it is the model underneath Grok Bot. Artificial Analysis ran benchmarks on it the same week. This is not a frontier leap. It is a reliability bump.
New model headlines do not change what your inbox looks like on Monday. What changes Monday is whether something is answering your missed calls.
The EU watermarking story is not new, and that is the part nobody said
I saw it covered this week as a fresh rule. It is not. The transparency duty comes from Article 50 of the EU AI Act, Regulation 2024/1689, which is 2024 law. What changed is attention, not the statute. I could not find a reliable this-week source for a new Commission action, so I am not going to invent one.
That correction is worth more to you than the headline was. A lot of AI news is old rules getting rediscovered by people who did not read them the first time. The practical piece stands either way. If you publish AI written copy, AI made ads, or AI generated product images into Europe, unlabeled output is a compliance problem instead of a style choice. Put a disclosure line in now. It is cheaper than a takedown.
DeepSeek is a pricing story, not a product story
V4 Pro is being read as price and benchmark pressure on the Western labs rather than a feature list. There is a V4 Pro PDF hosted on Hugging Face. I could not confirm an official this-week checkpoint from DeepSeek, so treat the timing loosely.
Someone is going to ask you why you do not just use DeepSeek since it is cheaper. Fair question. My answer is that I already run a capable model at zero marginal cost, so a price cut on tokens does not move me. Tokens were never the expensive part. The expensive part is somebody building the system, keeping it running, and answering for it when it gets something wrong. I do not resell tokens. I sell the installed system.
That is the week. One real product launch, one incremental model, one old law wearing a new headline, one price war I am not in.
The model in the news is the one already running my company. If you want that pointed at yours, that is what we do.
Sources
- xAI: Introducing Grok Bot
- Grok Bot product page
- Grok Bot pricing discussion on Hacker News
- xAI: Grok 4.6 release notes
- Artificial Analysis: Grok 4.6 benchmarks and analysis
- EU AI Act, Regulation 2024/1689 (EUR-Lex)
- EU AI Act, Article 50 text
- European Commission: regulatory framework for AI
- DeepSeek V4 Pro technical PDF on Hugging Face
- The podcast episode that seeded this roundup (Spotify)