// this week in AI

This week in AI: the agent moves into WhatsApp

2026 · 09 · 11 — THIS WEEK IN AIThe agent moves into WhatsApp

Five stories, and for once they all share the same thread. This week artificial intelligence stopped answering questions and started doing things on someone's behalf: buying, booking, writing emails, answering Google searches without sending anyone to the site that had the answer.

They're all verifiable facts, with the source at the bottom of the page. Below you'll find what happened and — above all — what changes for freelancers and small businesses that don't have an IT department to hand the question to.

1 · Meta puts an agent inside WhatsApp

On September 8, Meta unveiled Muse, its personal agent. The difference from any ordinary assistant comes down to a single verb: Muse doesn't just answer, it acts. It sends emails, fills out forms, books trips, completes purchases with a card generated for that single transaction. It runs on a dedicated virtual machine with a second agent that controls what it can reach on the network, asks for approval before sensitive actions and keeps a log of everything it does.

You use it from a dedicated app, from the muse.ai website and — this is the point — inside WhatsApp. For now it's available only in the United States and only to adults, free for normal use with paid plans for those who want more. It's not in Europe yet, and with the transparency obligations in force since August, it will probably arrive later and with a few more restrictions.

2 · OpenAI releases GPT-6 Astra

On September 3, OpenAI released GPT-6 Astra, its new flagship model: first to a small group of organizations, then rolling out to ChatGPT Plus, Pro, Business and Enterprise and via API. It's built for long, multi-step jobs — using a computer, browsing, writing code, seeing a piece of research through — more than for the quick one-off question.

No pleasant surprises on price: pay-as-you-go pricing stays at $10 per million input tokens and $50 per million output tokens, in line with Anthropic's flagship model. The announcement came in the middle of the security controversy that followed July's attack on Hugging Face, and OpenAI itself said it added safeguards before the release.

3 · Mistral raises €3 billion and promises compute in Europe

The French lab Mistral has closed a €3 billion funding round at a valuation above €21 billion: the largest round ever completed by a European tech company. It's led by Samsung Electronics and the Scaleup Europe Fund managed by EQT and PSG Equity, with Nvidia, BlackRock and — as a new investor — Luxembourg also taking part.

The part that matters isn't the figure, it's where it's going. Mistral says it wants to build up to 1 GW of computing capacity in Europe by 2030, with regional endpoints and the option for customers to choose which region processes their requests. It's not aiming to be “the European ChatGPT” for the general public: it's aiming at governments and businesses that want to use AI while keeping control over where their data ends up.

4 · NSA, CISA and FBI: know which model runs underneath

In joint advisory AA26-251A, published on September 8, the three US agencies claim that six Chinese companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI — extracted billions of tokens from Western frontier models to train their own, using fraudulent accounts, premium subscriptions bought in bulk and routing services to get around regional blocks. According to the advisory, this practice is the core of their development strategy, not a sideline.

It's an accusation, not a verdict, and the companies involved haven't acknowledged it. But the practical consequence is already here: around those models — cheap and often open — an undergrowth of tools has sprung up that nobody declares on an invoice, and they've ended up inside the workflows of quite a few small businesses without anyone noticing.

5 · Google's related questions are now answered by AI

Two independent analyses, released on September 9, say the same thing. The “People also ask” box — the one that opens accordion-style below the results — is now written by Google's artificial intelligence. AlsoAsked, on a set of almost 20 million searches, measures 97% of answers generated by AI Overview in the first week of September, up from 86% in August. Another survey puts it at 100% since August. Fourteen months ago the share was 12%.

That box was the side door through which a well-written page picked up visits without having to rank first: all it took was a good answer to a side question. Now Google gives the answer inside its own page, and the click never comes. The pages that keep bringing in work are the ones that answer a precise commercial question — who does it, how much it costs, how to get it done nearby — not the ones explaining a generic concept that the AI sums up in three lines.

In short

The week's thread is a single one: AI has gone from answering to doing. Meta put an agent inside WhatsApp that buys and books on its own — for now only in the United States, but it's the channel where real work gets done in Italy. OpenAI released GPT-6 Astra, and anyone already paying for ChatGPT gets it with no price increase. Mistral raised €3 billion to build compute in Europe, and with it the option to choose where your data ends up. The US agencies put in writing that behind certain cheap tools there's a murky story, and the question “which model runs underneath?” has become one any serious provider should expect. And Google has closed the last side door to free traffic: from here on, the only pages that bring visits are the ones answering a question that leads to a quote.

Sources: Meta (Muse announcement) · PBS News (what Muse does and where it's available) · Al Jazeera (GPT-6 Astra and the safety questions) · OpenAI (official GPT-6 Astra announcement) · TechCrunch (Mistral's €3 billion) · CISA (joint advisory AA26-251A) · Search Engine Roundtable (related questions written by AI).

Frequently asked questions