Instagram just started guessing how old your customers are, and it is doing it everywhere. On November 9 Meta took its AI age detection global: if the system thinks an account belongs to a 13 to 17 year old, that account moves into a locked-down teen account, no matter what birthday was typed at sign-up. If your shop sells sneakers, phone cases, makeup or anything else teenagers buy with their pocket money, this one lands on your page. The same day brought a big new model from Anthropic, a dirt-cheap agent model from Meta, and a startup that lets AI run real businesses. Here's what matters and what to do about it.
Instagram's AI now decides who counts as a teen
Meta has used AI age estimation in the US, the EU and Brazil for a while. Now it is rolling out worldwide on Instagram, with Facebook getting wider use and Threads due later this year. The system estimates age from signals on the account, and anyone it puts at 13 to 17 lands in a teen account: private by default, daily time limits, parental controls, and strangers can't message them.
That last part is the one to read twice. A shop page is a stranger to most of its customers. Meta's announcement doesn't spell out exactly how teen accounts treat business pages, so plan for the cautious version: don't count on starting a DM with a young customer. Cold outreach, "we saw you liked our post" messages, follow-ups to people who never wrote to you first: expect some of that to stop reaching anyone.
What still works is the conversation the customer starts. When a teenager writes to you, the clock is running and they are already interested. Answer fast, answer well, and close in that same thread. If you lean on automated welcome messages, our piece on welcome DMs that don't feel like spam is worth a reread with this change in mind.
One more reason Meta is moving now: in August it settled with 52 US attorneys general over harms to young users, and stronger age checks were part of the promise. This won't be the last tightening.
Anthropic ships Claude Fable 5.2 and Mythos 5.2
Anthropic released upgraded versions of its two top models. Fable 5.2 is open to everyone, with extra safeguards around cybersecurity and biology. Mythos 5.2 is the same model with those capabilities left in, available only to vetted security and life-science organizations. Anthropic says outputs are up to 30% faster than the 5.1 versions, and it claims the model helped solve two open problems in black-hole physics and designed three electronic devices that worked first time.
The price tells you who it's for: $8 per million input tokens and $40 per million output. That is heavy-duty pricing for coding and research teams, not for answering "is this in stock in size 38?"
So why should a seller care? Because last year's frontier becomes next year's default. Every jump at the top pushes the cheaper tiers up a notch, and the tools you use for captions, product descriptions and customer replies quietly get smarter. You don't need to buy the expensive model. You just get to enjoy what it drags along behind it.
Meta's Muse Ember: a capable model for ten cents
At the other end of the price scale, Meta launched Muse Ember, the smallest model in its Muse family. It is built for the boring, repetitive work inside AI agents: calling tools, routing requests, pulling details out of documents. The price is $0.10 per million input tokens and $0.50 per million output for normal-length prompts, it reads images, audio and video, and it has a one-million-token context window. Meta claims it beats rival small models on "intelligence per dollar"; those are Meta's own benchmarks, so treat them as a sales pitch until others test it.
Why it matters: the expensive part of a good AI assistant is rarely the one clever answer. It's the hundred small steps behind it: checking stock, reading an order, deciding which question this is. When those steps cost almost nothing, an assistant that actually checks things before it replies becomes affordable for a small shop, not just for big brands. That is the gap between a keyword bot and a real AI assistant, and it keeps getting cheaper to close.
A startup let AI agents run three real businesses
San Francisco startup Skyfall AI launched Steve-001, a team of AI agents that runs small online businesses the company bought. The numbers it shared are mixed, which makes them more believable. A social media API tool went from $3,649 to $8,307 in monthly revenue between July and October. A Shopify referral app is sitting about 3% below where it was when they bought it. A third business is too new to judge.
The interesting part isn't the revenue. It's the guardrails. A human owner sets direction and approves the big decisions, and every action passes through rules that sort it into four buckets: do it, do it and tell me, wait for my approval, or stop. Refunds of $30 or less can go through on their own; bigger ones wait for a person.
That is exactly the right mental model for any seller handing work to AI. Decide in advance what it can do alone and what it must pass to you. We wrote down the rules we'd use in when AI should hand off to a human. Also note what drove the growth: answering signup questions properly. Unanswered questions are lost money, in a SaaS tool and in your DMs alike.
Adobe's creator survey: AI helps, but people settle
Ahead of its MAX conference, Adobe published a survey of more than 2,000 US creators run with The Harris Poll. 78% of creative AI users said AI sped up their business or audience growth. But 68% admitted they sometimes accept a result that doesn't quite match what they wanted, and 47% had abandoned a project halfway because getting the right result took too much effort. And 77% said the final call should always stay with a human.
The same day, Disney signed a multi-year deal with Adobe to build custom Firefly models trained on Disney's own characters and styles. Big brands want AI that stays on-brand. You should want the same. If a generated product photo or caption is "almost right," it's wrong. Your customers notice the shirt that's a slightly different blue.
Quick takes
- Apple starts charging for extra AI. Apple launched $49.99 one-time "intelligence usage packs" that kick in after you hit your monthly AI limit. Only Logic Pro has them so far, with Final Cut Pro, Pixelmator Pro and iWork expected next. Free AI on your phone now has a meter.
- ChatGPT Enterprise lands on AWS Marketplace. Companies can now buy OpenAI's business plan through their existing Amazon cloud bill. Mostly a big-company story, but it shows where the AI money is flowing: into the plumbing.
- An AI host got breached. Forbes reports that AI cloud provider Fireworks AI disclosed attacks that exposed customer data and parts of unfinished models, and some customers have moved work elsewhere. If a tool holds your customer chats, ask where they're stored and who can see them.
- Anthropic's IPO numbers, walked back. Reuters reported that Anthropic sent bankers IPO details, including a raise of up to $69.3 billion, then told them the pricing information was inaccurate. Big money, still settling.
Today's thread runs through every story: AI is getting cheaper and stronger, and the platforms are getting stricter about who you can reach. That makes the conversations customers start with you more valuable than ever. If you want every one of those answered quickly and correctly on Instagram, WhatsApp and Telegram, see how Vardast handles it.