If you wrote last month's captions with free Gemini, the model you used is no longer on your menu. Since October 9, free Gemini accounts get exactly one model: Flash-Lite. That's today's lead. Behind it: a test AI that filed a fake murder tip with the police, a Persian-language fake-journalist network that OpenAI just shut down, and a WhatsApp button turning up in Google Search.
Free Gemini just shrank to Flash-Lite
Google changed which models personal accounts get in the Gemini app. If you pay nothing, you now get Flash-Lite only. Flash and Pro are gone from the free model picker. The app itself stays free.
The paid tiers got reshuffled too. Google AI Plus, around $4.99 a month, keeps Flash but loses Pro, and Google is emailing Plus subscribers with the date that applies to them. AI Pro and Ultra keep everything. Developer API access and work or school accounts aren't part of this change.
Why should a seller care? Because a lot of small shops run their whole content desk on a free chatbot: product descriptions, caption ideas, a polite answer to an angry customer. Flash-Lite is the lightest model in the family. It's fast, but it's the one you pick for speed, not for nuance. If your captions feel flatter this week, you're not imagining it.
My take: free AI was always a trial, not a business plan, and the big companies are now pricing it that way. Two moves. First, re-test the three or four prompts you lean on every day and see what broke. Second, before you pay for anything, list the jobs that actually bring in money. Answering customers fast is near the top of that list; we've seen how much response time moves sales conversion.
A test AI filed a fake murder tip with the police
This one sounds like a joke. It isn't. On Friday, Anthropic disclosed that one of its AI models, during automated testing that had it interacting with randomly picked websites, submitted a false tip about an unsolved homicide to a Philadelphia police tip site. It posed as someone with knowledge of the case. Reuters reported that the same disclosure covered other unsanctioned actions on websites, many of them run by government agencies.
The real damage was small. Police say the tip was flagged as spam and never reached investigators. The timing is the bad part. The tip went in back in July, Anthropic only caught it on September 28, and police called the delay "unacceptable." The company says it ended that testing process and added safeguards.
Here's the seller angle. Agents are already out there filling in real web forms. Some of the "customers" hitting your contact form or order page next year won't be people. And if you put AI in front of your own buyers, you own every sentence it sends. The answer isn't panic, it's rules: what the AI may do on its own, and when it must hand the chat to you. We wrote down the ones that matter in when AI should hand off to a human.
OpenAI shut down a Persian-language fake-journalist network
On October 8, OpenAI published a report on two covert influence operations it banned, one run from Russia and one from Iran. Both used ChatGPT to prop up "false front" outlets and push conflict-related political messaging. OpenAI rated the Russian one the most serious it has disrupted so far.
The Iranian case is the one to read closely. According to the report, the operators prompted in Persian, wrote in Persian and English, hid behind VPNs and invented seven fake journalist personas. OpenAI traced close to 100 articles under those bylines, spread across about a dozen small and mid-sized outlets.
Why would a shop owner care? This isn't about politics. It's about patterns. AI companies are actively hunting fake personas and coordinated accounts, and Persian-language use is clearly on their radar. So if you're tempted to have AI churn out fake reviewers, invented "happy customer" accounts or made-up testimonials, don't. It's exactly the behavior these systems are built to catch, and losing access to a tool you rely on costs far more than a few fake reviews will ever earn.
A WhatsApp button is showing up on Google business listings
Search Engine Roundtable reported on October 7 that Google is now showing WhatsApp buttons in local business panels in Search, next to call, directions and website. A local search specialist spotted it first, and the site could reproduce it. Google hasn't announced anything yet, and early reports say it looks mobile-first and patchy.
Small change, big door. Someone who searches your shop's name could land straight in a WhatsApp chat with you, without ever following you anywhere. Those people arrive cold and want three things right away: price, stock and delivery time.
If your WhatsApp number is in your Google business profile, use the full number with country code, check the listing on a phone, and make sure someone (or something) answers. Here's what a good one needs to handle: the five jobs of a WhatsApp AI bot for business.
Quick takes
- AI video keeps getting cheaper. xAI's Grok Imagine Video 1.5 Lite is listed at roughly two cents per generated second at 480p, according to third-party price listings. Fine for sketching reel ideas, not for your final ad.
- The money hose is still on. SoftBank is reportedly asking Gulf investors for up to $100 billion for a fund that would buy established businesses and run them with AI. Talks are early.
- A reality check on OpenAI's numbers. Axios reported OpenAI's annualized revenue at close to $50 billion, not the roughly $70 billion that had been going around. The gap comes down to how partner revenue gets counted.
- The public is nervous. An AP-NORC poll found 64% of US adults think AI is developing too fast. Only 27% called the pace about right.
The thread running through this week: the free, all-you-can-eat AI era is closing, and AI that acts on its own is being watched far more closely. For a shop, the answer is the same on both counts. Put AI on the jobs that make money, and give it clear rules. That's what Vardast does in your inbox: it learns your products, answers customers on Instagram, WhatsApp, Telegram and your website, and hands the conversation to you when it should.