This week an AI assistant learned to pay. Meta's new agent app, Muse, can now check out on Shopify stores and across PayPal merchants without the shopper ever opening a store page. Your next "customer" may be a piece of software sent by a customer. That's the story I can't stop thinking about, and it turns out the rest of the week's news circles the same idea.
Below: the Muse deal, the plumbing Stripe is building so agents can pay cleanly, six big banks saying "slow down", a price cut from OpenAI that makes AI cheaper for everyone (your competitors included), and a first for the UN. I read it all so you don't have to.
Meta's Muse turns from chatbot into checkout
Muse launched in the US on September 8. It's a personal agent that connects to your email, calendar and payment accounts and then does things for you in its own browser: books, fills forms, buys. It pulled roughly 730,000 downloads in its first five days and briefly sat above ChatGPT as the top free app on Apple's US store.
Now the part that matters for anyone who sells. Meta opened Muse to outside "connectors", and the first big names signed on:
- Shopify says it is partnering with Muse on agentic checkout with Shop Pay across Shopify stores.
- PayPal says its customers will be able to shop and check out through Muse across PayPal merchants worldwide.
- Instacart turns a prompt like "Taco Tuesday" into a ready-to-pay grocery cart.
- Expedia lets Muse plan trips and book hotels inside the app.
Amazon, meanwhile, blocked Muse from its site, citing privacy and credential concerns. Tells you how seriously the big stores take this.
Why it matters for sellers: the shopper who asks an assistant "find me a black linen shirt under $40 and buy it" doesn't see your beautiful feed. The agent sees your catalog: title, price, stock, photos, shipping. Merchant guides on the Shopify side already say products need a clear title, at least one image and a real price to qualify, and that bundles and custom items are left out. Direct checkout is US-only for now, and it won't reach Iranian stores any time soon. But the habit it teaches ("let the assistant handle it") travels fast. If your product info lives in your head and in old captions, start writing it down properly. A clean product catalog your AI can actually read is the same asset whether the reader is your own assistant or someone else's agent.
Stripe gives shopping agents a proper door
Agents that shop today mostly do it the clumsy way: read the page, guess which button to press, type into fields. Stripe is testing a cleaner route called WebMCP. The payment page itself tells the agent what it can do (read the amount due, pick a payment method, fill a field, submit) as structured tools.
Stripe's own benchmark: with WebMCP, agents used 42% fewer tokens and 38% fewer tool calls to finish a payment. Stripe labels it experimental and built on a proposed web standard, so treat it as an early sign, not a finished road.
Why it matters: the web is quietly growing a second front door, one for people and one for machines. Stores that make the machine door easy will get the agent traffic. Everyone else gets a confused bot that gives up at the shipping field.
Six banks to AI agents: not so fast
The same week the agents got wallets, ASB, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest published a set of principles for agent shopping. Their worry list is long: fraud, scams, disputes, agents that buy the wrong thing or spend too much, and agents that nudge people toward payment methods with weaker protection.
They plan to take proposals to regulators, including a rule that shoppers must be told when an AI agent is part of a transaction, plus clearer insight into how agents make their choices.
Why it matters: honestly, I'm glad someone said it. For a seller, the risk is simple and practical: more orders you didn't expect, more "I didn't mean to buy that" refunds, more disputes. Have your return and refund rules written clearly before the agents arrive. Our piece on handling returns and refunds in chat covers the human side of that moment, which doesn't go away because a bot placed the order.
OpenAI halves its prices with GPT-6 Sol and Luna
OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, a few weeks after the bigger GPT-6 Astra. The headline is price. Sol costs $2 per million input tokens and $10 per million output, half of what its predecessor cost. Luna, the small one, costs $0.10 and $0.50. OpenAI says these are permanent prices, not a launch promo.
Both are rolling out in ChatGPT for paid plans, and Free and Go users get Luna in the desktop app.
Why it matters: you won't buy tokens yourself. But every app you use that has "AI inside" (your design tool, your inbox, your competitor's reply bot) just got cheaper to run. Cheaper AI means more of it everywhere, which means customers will expect fast, smart replies as the normal, not the bonus. The bar is going up, whether you like it or not.
The UN Security Council sits down to talk about AI risk
Today, September 23, France is chairing the Security Council's first meeting focused specifically on the safety risks of increasingly capable AI, including losing human control of it. The expected speakers include OpenAI's Sam Altman, Anthropic's Dario Amodei, Hugging Face's Clément Delangue and scientist Yoshua Bengio.
Nothing here changes your Instagram shop tomorrow. I include it because the rules that eventually come out of rooms like this (disclosure, labeling, who is responsible when an agent gets it wrong) are exactly the rules that will decide how AI is allowed to talk to your customers.
Quick takes
- Anthropic released Claude Opus 5.5, priced about 20% below its previous top model. Same pattern as OpenAI: stronger and cheaper at once.
- Google folded Colab into its Google AI plans, with priority computing for Pro and Ultra subscribers. Mostly a builder story, but it shows the bundling war isn't slowing down.
- OpenAI DevDay is on September 29. Expect more agent and developer news next week.
- From Monday's digest, if you missed it: Meta One put a price on AI customer replies inside Instagram and WhatsApp.
The thread through all of this: buying is becoming a conversation, sometimes with a person and increasingly with a machine acting for one. Whoever answers clearly, fast and with the right facts wins that conversation. That's the job Vardast does in your Instagram, WhatsApp and Telegram DMs today: it answers from your real catalog, day and night, and hands off to you when a human touch is needed.