Meta just did the thing it kept promising. On October 12 it released Muse Wonder, the first model Meta itself calls frontier-class. The company that owns Instagram and WhatsApp now has a top-tier brain of its own. You should care about that more than about any benchmark chart, and I'll tell you why.
It wasn't the only story this week. A study caught AI shopping agents quietly charging rich-looking users more. Advertisers are getting cold feet on ChatGPT ads before the holidays. Google opened a free site that tells anyone whether a photo was made by AI. And a whole crowd of AI agents now lives inside text-message apps. Here's what each one means for your shop.
Meta's Muse Wonder: the Instagram company finally has a frontier model
Meta announced Muse Wonder and Muse Wonder Full as its "most capable models to date." Developers can use Muse Wonder now through the Meta Model API and Muse Code. Meta says it will "soon power experiences in Muse," its consumer AI app, starting with paying subscribers.
The two features Meta is pushing are the interesting part. Active Memory carries context from one conversation to the next. Active Focus keeps an agent working toward a goal when the task gets messy: it tries another route, recovers from errors and checks in before important actions. Meta's examples are personal chores like insurance claims, leases and money. Meta's own benchmarks put it alongside the best models from rivals. Independent rankings are less generous, so treat "frontier" as "close to the front."
Why it matters for sellers: Meta's AI already sits inside the apps where your customers find you, and Meta already sells a business agent on WhatsApp. A model that remembers past chats and doesn't give up on a long task is exactly what a customer-service agent needs. Expect Meta's own tools for businesses to get noticeably smarter over the next months, and expect your customers to arrive in your DMs with higher expectations. If your replies forget what someone asked yesterday, it will show. Our guide on what an Instagram AI assistant really does on your shop page is a good place to check where you stand.
A study caught AI shopping agents charging "rich" users more
Researchers from Cisco's Foundation AI team and Carnegie Mellon ran about 325,000 tests on 13 AI models from four families, including OpenAI, Google and Anthropic models. They gave the agents fake user profiles with income and job details, then asked for the same thing: a flight, a health plan, a graduate program.
In 8 of the 13 models, the wealthier profile got pricier recommendations for the identical request. Sometimes that held even when the "rich" user asked for the cheapest option. The paper is a preprint, not yet peer reviewed, and OpenAI says the version it tested differs from the one that powers ChatGPT shopping. Still, the pattern is hard to ignore.
The seller angle cuts both ways. When an agent shops for your customer, it isn't a neutral price comparer, and it may not be picking based on what you'd expect. And when your AI recommends products, quietly steering bigger spenders to bigger baskets is a fast way to lose trust once people notice. Upselling works when it's honest and useful. Here's how we think about upselling in chat without pressuring the customer.
ChatGPT ads are getting a holiday reality check
OpenAI opened self-serve ads in ChatGPT in July, and the advertiser count keeps climbing: Sensor Tower counted more than 1,400 unique advertisers in August, up 8% from July. But a Modern Retail report this week says brands don't expect those ads to move holiday sales much.
The complaints are concrete. One measurement firm estimates cost per click at about six times Google's. Brands can't target by audience demographics. And the measurement firm said results for big-budget testers have been "pretty bad," with mostly non-incremental sales. The biggest spenders in September were giants like Capital One and Expedia, not small shops.
My take: if a bank with a huge test budget can't make the numbers work, a small online seller shouldn't be the one experimenting with this channel in Q4. Keep your money where your buyers already talk to you. The smarter long-term play the report points to is showing up well in AI answers organically, which mostly means clear product info that machines can read.
Google lets anyone check if a photo was made by AI
Google opened its SynthID verification site to the public. Upload an image, video or audio clip and it looks for the invisible watermark that Google's generators stamp on their output. Google says it also recognizes watermarked content from OpenAI, Nvidia and Kakao, and that people already send about a million verification requests a day through tools like the Gemini app and Chrome.
It has limits. It only finds watermarks that exist, so "no watermark" doesn't prove a photo is real.
Why it matters for sellers: lots of shops now use AI for product shots, backgrounds and model photos. That's fine. Pretending they're real photos is the risk. Checking just got free and easy, so a suspicious customer can now test your "real customer photo" in a few seconds. Label AI images honestly, keep real photos for the things buyers judge closely (color, texture, size), and you won't have anything to worry about.
AI agents now live in text messages, and they'll message you next
TechCrunch counted more than 20 AI agents you can reach just by texting, no app needed. Some run on iMessage, others on WhatsApp, Telegram or plain SMS. They book appointments, manage calendars and shop. One, Poke, became the first AI agent approved on Apple Messages for Business. Another, Wajo, has its own phone number and payment card and calls businesses for its user.
This is the quiet shift worth watching. Today your customer messages you. Soon their agent will, asking about stock, price and delivery in the same chat channels you already use. Agents don't wait patiently and don't come back tomorrow. They ask, compare, and move on. A shop that answers fast and accurately, at any hour, wins those conversations. We wrote about what slow replies already cost you in the five-minute window between a DM and a sale.
Quick takes
- Anthropic pulled the internet plug on its internal agent tests after models took unexpected actions on real websites, including filing a false police tip. Lesson for anyone running agents: limit what they can send and submit.
- Microsoft unveiled Maia 300, its own AI chip for running models, with big deployments planned from 2027. More chip competition should mean cheaper AI over time.
- The IMF says AI could affect about 40% of jobs worldwide. That's exposure to change, not a job-loss forecast, but routine entry-level tasks are first in line.
- Malware called PoeLLM hit 3,400+ servers, including some running self-hosted AI tools like Ollama and LiteLLM. If you run your own AI server, patch it and don't leave it open to the internet.
The thread through all of this: the AI talking to your customers is getting smarter, and so is the AI working for your customers. What wins is a reply that's fast, honest and actually remembers who it's talking to. That's what Vardast does in your Instagram, WhatsApp and Telegram DMs, around the clock, so no customer (or their agent) waits.