AI Next WaveOctober 7, 2026
Free Gemini Shrinks Friday. The Strongest AI Now Has Gatekeepers.
Google and OpenAI are changing what free users get, and Anthropic is sorting who gets its cyber tools. South Korea's bank breaches show why those gates matter.
The Short Version
On Friday, anyone using Google's Gemini app without a subscription gets moved to Flash-Lite, the company's lightest model. Flash will take a plan that costs about $5 a month, and Pro will take one that costs about $20. OpenAI picked a different lever on Monday. It kept the free plan and said it will start testing visual ads inside free ChatGPT image generation later this month.
On Tuesday, Anthropic rolled out three tiers that decide which security professionals get its most advanced cyber abilities, while its regular models keep safeguards that block most cyber work. That move lands beside a warning from South Korea, where investigators found traces of an open-source attack kit, built to plug into models from Anthropic, OpenAI and DeepSeek, in a wave of breaches at seven banks and lenders.
Money decides what you get from Gemini. Permission decides what you get from Anthropic's cyber tools. My read is that both gates will shape your week: check what your free plan still does after Friday, and check the side doors on anything you run online.
What Free Gemini Users Lose on Friday
Google's help page spells out the new line-up for personal accounts. Starting October 9, people without a subscription get Flash-Lite only. AI Plus gets Flash-Lite and Flash, and Google says Plus subscribers will get an email telling them when their change takes effect. AI Pro adds Pro, and AI Ultra gets every model. Engadget, which lists AI Plus at $5 a month in the US and AI Pro at $20, points out that Plus subscribers lose Pro too.
Limits change shape as well. Google describes them as compute-based, refreshing every five hours until you hit a weekly cap, with Plus getting twice the standard allowance and Pro four times. You can now pick low, medium or high effort, and Android Authority notes that higher effort drains your allowance faster. Media generation, Deep Research, the Pro model and extended thinking all use more of it.
Google built Pro for advanced reasoning, complex math and coding, according to Engadget's account of the company's own description. Engadget calls Flash-Lite Google's most cost-effective model. I haven't found an independent test comparing the two on everyday work, so I'd judge Flash-Lite by your own tasks. For a meeting summary or a polite email, you may see little difference. For someone who has leaned on free Pro to fix a spreadsheet formula, debug a script or read a long contract, Friday is a pay-or-switch decision.
The help page doesn't say why Google made the change. As of Monday, Notebookcheck reported, Google's own plans page still showed the old line-up, so check the help page before you decide.
OpenAI Keeps the Model and Sells the Wait
OpenAI is charging free users in attention. Its Monday post says a new visual ad format will appear during image generation for people on the Free and Go plans, with testing starting later this month in the US with a first group of advertisers. The company says the ads will be clearly labeled and kept separate from the image you're making, and that advertising "does not influence the answers ChatGPT provides."
The Go plan deserves a second look. It costs $8 a month in the US, and OpenAI's help center lists it with Free as a plan where ads may appear. Plus, Pro, Business, Enterprise and Edu accounts are ad-free, and OpenAI says it doesn't show ads to accounts it identifies as belonging to people under 18. Free users can switch to an ad-free mode, but that mode comes with fewer messages and no image generation or deep research.
So the two companies are pricing the same free lunch differently: Google is trimming what the free model can do, while OpenAI keeps the model and sells the moment you spend waiting for a picture to finish. For anyone who uses these tools at work, the practical question is which cost you mind more: weaker answers, or an ad beside your client's draft graphic.
A Bookkeeper Runs the Numbers
Picture Rosa, a hypothetical bookkeeper who runs a one-person practice and has used free Gemini Pro to check spreadsheet formulas and draft client emails. Before Friday, she pulls up the last month of chats and saves 20 typical requests with the answers she actually used. That's her baseline. The numbers here are illustrative design choices, picked to keep the test small.
For two weeks after the switch, she runs the same kinds of requests on Flash-Lite. She counts an answer as accepted only if she used it with light edits, and she logs the minutes she spends checking or redoing work. Formula logic across several linked tabs and long PDFs are the likely failure points, and a heavy Monday could run into the five-hour limit.
The gain she's measuring is time, and the cost is cash. If Flash-Lite costs her an hour a week in rework, $20 a month for Pro pays for itself fast. If acceptance barely moves, she keeps her money. Rosa has the authority to stop the trial on day one if a wrong formula reaches a client, because a bad number on a tax worksheet costs more than any subscription.
Seven Korean Lenders and One Open-Source Kit
The Gemini change is about what AI you can use. South Korea's breaches show what happens when someone else uses it against an institution you trust.
Shinhan Bank disclosed the first breach on October 1, the Korea Herald reported, and by Sunday seven institutions had reported incidents: Shinhan, KB Kookmin, Hana and BNK Busan banks, Yegaram and Welcome savings banks, and Hyundai Capital. Shinhan put its count at about 25,700 customers. The Herald's tally across all of them reached about 66,000 people, and reports conflict on KB Kookmin's share. Exposed data included names, phone numbers, annual income and loan limits. The country's Financial Services Commission said Tuesday it had confirmed no cases of customers losing money, American Banker reported.
At the four banks American Banker detailed, the attackers came in through side doors: a loan-recruiter lookup service at Shinhan, an employee mobile work system at KB Kookmin, a sales-support system at Hana, and web pages at BNK Busan with weak checks on who was logged in. The trade paper described them as "what appear to be ordinary cyber hygiene failures rather than complex vulnerabilities."
That's where AI comes in. Investigators found traces of ARTEX, a Chinese-language penetration-testing tool posted openly on GitHub. Korea's SBS describes a planner agent that picks the attack direction and worker agents that run commands, with settings where users enter keys for Anthropic or OpenAI models, and a September update that added DeepSeek. An official at the Korea Financial Security Institute said the AI "did not act independently without human involvement." President Lee Jae Myung told his cabinet on Tuesday, Korea time, that signs had emerged of AI being used, but officials haven't said which model was connected or who the attackers are.
Kim Seung-joo, a Korea University professor, put it bluntly: "The problem is that vulnerabilities were not properly managed." I'd add that AI shrinks the time between a weak side door existing and someone finding it. Regulators gave the firms until Thursday to check every internet-facing system and fix the gaps.
The lesson travels. Days before the Korean attacks became public, Federal Reserve Vice Chair Michelle Bowman told US community bankers that defending against AI-powered threats "begins with strong cyber hygiene," including sign-in methods that resist phishing and tight controls on who can access what. American Banker notes those are the controls that guard against the gaps that showed up in Korea, a verification step that got bypassed and session checks that failed.
Anthropic Decides Who Gets the Sharp Tools
A tool like ARTEX borrows its reasoning from whatever model you connect, which puts some of the burden on the companies that sell those models. On Tuesday, Anthropic folded its Project Glasswing partnership into a three-tier Cyber Verification Program. Defense Access covers incident response, malware analysis and vulnerability work. Red Team Access adds authorized penetration testing, for organizations only. Specialized Access, with the fewest blocks, covers testing systems such as power grids, telecom networks and interbank transfer systems, reviewed with the US government.
Anthropic says its regular models keep conservative safeguards that block most cyber work, though anyone can still use them to review code, patch known issues, find flaws in their own code and sort security alerts. In the company's own test of multi-step attack scenarios, every task run without verified access was blocked on the first prompt. The company also says Glasswing partners found at least 129,000 verified software vulnerabilities between April and July, more than 33,000 of them rated critical or high.
Those gates apply to Anthropic's models. ARTEX's menu also lists OpenAI and DeepSeek, each with its own rules. I'd count a verification program as one supplier raising the cost of misuse. The Korean breaches suggest the bigger gap sat on the banks' side.
Opportunity Radar
Plenty of small lenders, mortgage brokers and accounting firms run partner portals and staff logins they've half forgotten. My hypothesis is that some would pay a fixed fee for a plain inventory of every internet-facing login, with a check on whether each one uses phishing-resistant sign-in. Test it by offering the inventory to five local firms at a set price. Walk away if most say their IT provider or cyber insurer already does it.
What You Can Do With This
If you use free Gemini or ChatGPT for work
Before Friday, save five prompts you rely on and the answers you got, then rerun them on Flash-Lite next week. Pay only if the gap costs you real time. If client graphics come out of free ChatGPT, expect ads beside them later this month, or move that work to an ad-free paid plan.
If you run a business with outside-facing logins
List every page a partner, contractor or employee can reach from the internet, including old portals. Turn on phishing-resistant sign-in where you can, and shut the pages nobody uses. Korea's attackers went through exactly those doors.
If you bank online
Ask your bank or credit union whether its partner and broker portals use the same sign-in protections as its customer app. Treat any caller who already knows your income or loan limit as a warning sign, since that's the kind of data the Korean attackers took.
The Bigger Picture
This week, the strongest AI started coming with two kinds of gates. One is a price tag: free users get the lightest model or an ad, and paying customers get the rest. The other is a credential: Anthropic now decides which security teams get its most capable hacking help. Both gates sort people at the lab. Korea's breaches show the limit of sorting there, because attackers apparently paired an open-source kit with ordinary gaps at the institutions themselves. My expectation is that more banks and businesses will learn their exposure from an attack like this one, and the ones that inventory their side doors first will have the shortest list of surprises.
References
Changes to Gemini model access and limits (Google Gemini Help, opened Oct 7, 2026)
Google removes free access to Gemini Flash and Pro on October 9 (Engadget, Oct 6, 2026)
Google is downgrading Gemini for users on budget plans (Android Authority, Oct 5, 2026)
Google Gemini drops Flash and Pro for free users on October 9 (Notebookcheck, Oct 5, 2026)
Building advertising for the way people use AI (OpenAI, Oct 5, 2026)
Ads in ChatGPT (OpenAI Help Center, opened Oct 7, 2026)
Introducing ChatGPT Go (OpenAI, Jan 16, 2026)
Police launch major probe as suspected AI hacks sweep through banks (The Korea Herald, Oct 6, 2026)
AI-linked hacks hit Korean banks through loan-agent sites (American Banker, Oct 6, 2026)
Anthropic expands its Cyber Verification Program into three tiers (Anthropic, Oct 6, 2026)