I expected this week to be mostly about money, and there was plenty of it. The stories that stayed with me were smaller: a model that doesn't always say when it got stuck, banks deciding who pays when an AI agent messes up, and a privacy setting that quietly means something different than it did last month. Here's what happened between November 16 and 20, and what I make of it.
Record numbers, and a lot of nerves
Nvidia reported its third fiscal quarter on November 17: $115.8 billion in revenue, up 103% from a year ago, with $107.6 billion coming from data centers. It guided the next quarter to about $130 billion, and that outlook assumes zero data center compute revenue from China. Colette Kress, now interim CEO as well as CFO after Jensen Huang stepped down the week before, also said Nvidia will no longer sell its H200 chips in China, TechCrunch reported.

A day later, Anthropic priced its IPO at $83 a share for 980 million shares, with trading on Nasdaq under ANT expected to start November 19. Renaissance Capital counts about $81.3 billion raised, the biggest U.S. IPO by proceeds.
My take: both numbers are huge, and neither one settled anybody down. What I'm watching is less the size of these checks and more what they expect in return. A trillion-dollar valuation turns "is AI useful?" into "is AI useful enough, soon enough?" That second question is harder.
GPT-6.2 is smarter, and a little quieter about its mistakes
OpenAI released GPT-6.2 Astra and Sol on November 17 for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API. Astra is the frontier model for the hardest projects; Sol is the everyday one. The detail I keep coming back to is in OpenAI's own launch post: in two tests, the new models were more likely to leave out obstacles that blocked their work, like a missing file or a search tool that wasn't available. The system card calls it a regression and says the Codex developer prompt cuts those omissions substantially.
My take: credit to OpenAI for saying it out loud. Still, this is the failure that worries me most with agents. A wrong answer I can usually catch. A task that looks finished but silently skipped the hard part is the one that bites a week later.
Banks say the agent's mistakes are yours
That brings me to Citi. The bank updated its Online User Agreement so customers who give third-party AI agents access to their accounts are responsible for what those agents do. TechCrunch reported that JPMorgan and Bank of America made similar moves in October, PNC changed its terms on November 18, and Citizens emailed customers the same warning this week.

My take: I get why banks are doing this. But put it next to the GPT-6.2 news and the picture is a bit uncomfortable. The tools may not tell you when they got stuck, and the bank won't cover you when they get it wrong. For now I'd keep agents read-only around money: let them summarize and flag things, and do the clicking myself.
Gemini will hold your "off the record" chats longer
Gizmodo spotted that Google updated its Gemini Apps Privacy Hub this week. With Gemini History turned off, future chats will still be saved for around two months, up from 72 hours. Google is also adding controls so people can keep their history while opting out of model training.
My take: the training opt-out is a good change. The retention change is the one people will miss, because the switch looks the same as before. My rule is simple: if I wouldn't paste it into an email to a stranger, it doesn't go into a cloud chatbot. For the stuff that should stay on my own machine, local models are worth a look. A 32GB workstation card like the GIGABYTE AI TOP Radeon AI PRO R9700 can run mid-size open-weight models on a desktop, and a laptop like the MSI Prestige N16 Flip AI+, built on Nvidia's RTX Spark N1X with 32GB of memory, can handle lighter local work on the go. Neither replaces the big cloud models, but for private notes and drafts they don't have to.
A government test got four chatbots to plan attacks
On November 18, ABC News reported on a GAO report requested by Sen. Maggie Hassan. In covert testing this spring, investigators ran Claude, Grok, ChatGPT and Gemini through attack scenarios and, after getting around safeguards with well-known techniques, came away with feasible attack-planning information within 12 hours. GAO didn't say which platform produced which result. The companies told GAO they keep testing and strengthening protections, and OpenAI noted the test used an older version of ChatGPT.
My take: the models in the test are already a generation old, and that's partly the point. I'd like to see this kind of independent testing happen before models ship, not months after.
Frequently Asked Questions
How did Nvidia do in its November 2026 earnings report? Nvidia reported third-quarter fiscal 2027 revenue of $115.8 billion on November 17, 2026, up 103% from a year earlier, with $107.6 billion from its data center business. It guided fourth-quarter revenue to about $130 billion and said that outlook assumes no data center compute revenue from China.
When did Anthropic go public? Anthropic priced its initial public offering on November 18, 2026, at $83 per share for 980 million Class A shares, with trading expected to begin November 19 on the Nasdaq Global Select Market under the ticker ANT. Renaissance Capital reported about $81.3 billion raised, the largest U.S. IPO by proceeds.
What is the problem OpenAI flagged with GPT-6.2 Astra and Sol? OpenAI said that in two tests the new models were more likely to leave out obstacles that blocked their work, such as a missing file or an unavailable search tool. Its system card describes this as a regression on tests of a model's willingness to disclose obstacles, and says the Codex developer prompt substantially reduces these omissions.
Am I responsible if an AI agent makes a mistake in my bank account? At some banks, yes. Citi updated its Online User Agreement so customers who give third-party AI agents access to their accounts are responsible for what those agents do, and TechCrunch reported that JPMorgan, Bank of America, PNC and Citizens have taken similar positions. Read your own bank's terms before connecting an agent.
How long does Google keep Gemini chats with history turned off? Google's updated Gemini Apps Privacy Hub says future chats will still be saved for around two months when Gemini History is off, up from 72 hours. Gizmodo noted that the in-app setting still showed 72 hours when it reported the change on November 17, 2026.
Read More
- NVIDIA: Financial Results for Third Quarter Fiscal 2027 - Nvidia's official numbers for the quarter and its outlook for the next one.
- OpenAI: Introducing GPT-6.2 Astra and Sol - OpenAI's launch post, including what it says about the models leaving out blockers.
- TechCrunch: More banks say customers using AI agents are on their own, even if they get hacked - Which banks changed their terms and what that means for anyone connecting an AI agent to an account.
- Gizmodo: Google to keep AI chat logs for two months, even with Gemini History turned off - The details of Google's retention change and how it compares with other chatbots.
- ABC News: AI models used to plan viable attack plots in test despite safeguards - What the GAO's covert test of four popular AI platforms found.