Two weeks ago I told you Google's new top model existed but wasn't for you yet. As of October 13, it is. Gemini 4 Argon is now open to paying Gemini users and to businesses building on Google's tools, and it arrived with a feature that matters more than the model itself: a mode where Gemini keeps working after you close the laptop.
On the same day, Shopify opened its billion-product catalog to almost any AI agent that wants to shop on someone's behalf. And Anthropic's own security team published a warning about small AI models copying themselves onto cheap gadgets, the kind of no-name TV box you might have in your shop right now. Three stories, one direction: AI is moving from answering questions to doing things. Here's what each one means for you.
Gemini 4 Argon is finally open, and it brought a "keep working" mode
Google first announced Argon on September 30 but gave it only to a small group of cybersecurity partners. On October 13 it went broad. Developers and companies can use it through Google's AI building tools, and Google AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Search. Google calls it its most capable model for complex coding, agentic, math and science work. Reuters framed the launch as Google trying to win back ground from OpenAI and Anthropic after months of delays and a leadership shakeup at DeepMind.
The app update is the part to watch. Gemini now has a tasks toggle next to normal chat. In task mode it keeps working while you're away from your computer, across your Google apps. It also connects to more than 40 outside apps, including Instacart, OpenTable, Adobe and Shopify. Google says Argon answers more directly, hedges less, repeats itself less and avoids clichés. I'll believe the cliché part when I see it, but it's the right goal.
For developers there's a clock running. Launch pricing starts at $2 per million input tokens and $10 per million output tokens, and Google says those rates double after November 30. That's a classic "come try us" discount.
Why it matters for sellers: the big three AI companies now all sell some version of "give it a goal and walk away." That's where your customers' expectations are heading too. If an assistant can book a dinner table without supervision, a buyer will expect your shop to confirm stock, price and delivery in one reply, not after three follow-ups. The model race matters less to you than this: fast, complete answers are becoming the baseline, not the bonus.
Shopify opens a billion products to any shopping agent
On October 13, Shopify said shoppers can now connect their own AI agents to its catalog of more than a billion listings, including products from stores that aren't on Shopify. Before this, each agent needed its own partnership deal, like the one Shopify signed with Meta's Muse agent in September. Now any compatible agent can plug in through a standard connector Shopify calls Shop MCP, search products, build a cart and pay with Shop Pay inside the chat.
Shopify is selling it as a personal shopper with rules. The shopper can set a budget limit or tell the agent to show only items with fast shipping. Every purchase still needs the shopper's approval. Merchants get controls too: they can manage catalog access and direct checkout for each AI channel. One detail worth knowing: opting out of the catalog doesn't stop ordinary web crawling, so your product pages are still readable to machines either way.
Why it matters for sellers: most of our readers don't sell on Shopify, so you won't flip a switch here tomorrow. But the lesson travels. When an agent compares products, it doesn't see your nice photos or your charming captions first. It reads names, prices, sizes, stock and shipping times. Messy or missing product data means the agent quietly picks someone else. The same is true for your own AI assistant in the DMs, which is why we keep banging on about training your AI assistant on a real product catalog instead of letting it guess.
My take: the store of the next few years has two front doors. One for people, one for agents. Most small shops have only built the first.
Anthropic warns small AI models may soon copy themselves onto cheap gadgets
Anthropic's Frontier Red Team published research arguing that small AI models may be months away from being able to copy themselves onto ordinary internet-connected devices: smart ovens, off-brand Android TV boxes and similar hardware. In their tests, a 9-billion-parameter model got itself running on 8 of the 9 hardware types they tried. They also ran a larger open model spread across 15 cheap TV boxes, about $55 each. It worked, slowly.
The important caveat: this is a trend projection, not a working attack. Smaller models can already do many of the individual steps but still stumble when chaining them together. The researchers' worry is what happens when they stop stumbling. A self-copying model with no central server would be hard to shut down, and unlike normal malware it could figure out unfamiliar devices on its own.
Why it matters for sellers: lots of small shops run on cheap, never-updated gadgets. The security camera recorder, the TV box playing product videos in the store, the old router nobody has logged into since it was installed. None of this is an emergency today. It's a nudge to do the boring things now:
- Change default passwords on every connected device.
- Update firmware or replace gadgets that never get updates.
- Keep your payment and order systems on a separate network from the TV box and the guest Wi-Fi.
Quick takes
- Gemini Enterprise turns Argon on by default from October 27. Business admins can already switch it on in the web and mobile apps, but the toggle starts as off until then. If your team uses Gemini at work, expect the model under it to change that week.
- Cisco put AI agents into Webex meetings and calls. Employees can invite agents to work on multi-step jobs across calendars and documents, with human oversight and audit trails. Meetings are slowly becoming places where software takes notes and also does the follow-up.
- Ambi raised $26 million for a meeting agent that can optionally listen all the time on phone, Mac and Apple Watch. Useful, maybe. If you use something like this in a shop, tell customers before it hears them.
- DigitalOcean launched preview tools to test and monitor AI agents: traces, failure flags, evaluations and simulations. Good sign. Checking what an agent actually did is becoming a product category, not an afterthought.
Here's the week in one line: agents are learning to shop, to keep working while nobody watches, and in the worst case to spread on their own. For a seller, the safe move is the same one it has always been, just more urgent. Have clean product data, answer customers fast and completely, and keep a human in charge of anything that touches money. If you're doing the math on handling all those messages yourself, our breakdown of customer support automation cost is worth five minutes, and so is our note on what a website chatbot needs to actually sell. When you're ready to have every DM answered with your real prices and stock, Vardast does that on Instagram, WhatsApp, Telegram and your website.