
A busy week, and not all of it good. Google and Anthropic both shipped new top models and ChatGPT started paying for things. At the same time Nvidia’s CEO resigned, private AI chats leaked through a vendor, and an AI coding tool was found saving passwords into public code. Prices are per million tokens and come from the companies themselves.
Gemini gets a faster everyday model
Google released Gemini 4 Hydrogen on November 12. It replaces Gemini 3.8 Flash as the model most people meet in the Gemini app and in Search. Google says it finishes tasks 2.5 times faster and makes things up more than 5 times less often. Free users get it automatically. For developers it costs $1.50 in and $7.50 out.
Why it matters: if you use Gemini for free, it should get quicker and more accurate without you doing anything.
Google’s strongest model: Gemini 4 Carbon
On November 10 Google launched Gemini 4 Carbon, built for long, hard jobs like debugging, research and multi-step agent work. It is in the Gemini app and AI Mode in Search for Google AI Ultra subscribers, and in the API for paying developers. The launch price is $2 in and $10 out. It doubles after November 30.
Why it matters: Google’s best model is now in a consumer app, not only behind an API.
Claude Fable 5.2 and Mythos 5.2
Anthropic released Claude Fable 5.2 on November 9. It is aimed at hard coding, finance and science work. Mythos 5.2 is the same model with fewer restrictions, offered only to vetted organizations. Fable 5.2 costs $8 in and $40 out. Pro and Team subscribers can use it too, up to half of their weekly usage limit. Anthropic still suggests most people start with Claude Opus 5.5.
Why it matters: the most capable Claude is no longer only for big companies.
Claude Sheets, Whiteboards and Data
Also on November 10, Claude got three new tools on all paid plans. Sheets (beta) builds and edits spreadsheets with you and tracks every change. Whiteboards (beta) is a shared board for brainstorming that Claude can sort out and turn into a document or slides. Data, which used to be called Dashboards, is now fully released and works with your connected data.
Why it matters: AI assistants are moving out of the chat box and into the tools people already work in.
Claude joins your video calls
Claude Meetings (beta) lets Claude join a Google Meet call as a visible participant. You can ask it questions out loud or in the chat, and it can draft documents during the call and write notes and action items afterwards. It keeps a transcript and a summary, not the audio. It is rolling out to Team and Enterprise plans first.
Why it matters: the meeting note-taker can now do some of the follow-up work too.
ChatGPT agents can now pay for you
On November 13 OpenAI started testing ChatGPT Wallet. ChatGPT’s agents can finish a purchase with a saved card after you check the items, the total and the card, and approve with Face ID or a fingerprint. Neither the agent nor the store sees your real card number. For now it is only open to some ChatGPT Pro subscribers in the US, and spending limits haven’t arrived yet.
Why it matters: letting an AI spend your money is a big step, so watch which safeguards are there and which are still missing.
Businesses get profiles inside ChatGPT
OpenAI also opened a beta on November 10 for verified business profiles in ChatGPT, a bit like a Google Business Profile. A business can keep its own details up to date, and some can add an agent that answers customer questions.
Why it matters: when you ask ChatGPT about a shop or a product, more of the answer may come from the business itself.
Meta’s cheapest model: Muse Ember 1.0
Meta released Muse Ember 1.0 on November 9, its smallest and cheapest model, made for high-volume work like tool calls and long documents. It costs about $0.10 in and $0.50 out, and it is also in preview on Google Cloud.
Why it matters: cheap, capable models make it realistic for small apps, like the ones I build, to include AI.
Google Cloud model page · TestingCatalog report
Nvidia’s CEO resigns
Nvidia, the company that makes the chips most AI runs on, announced on November 12 that Jensen Huang resigned as CEO and left the board, effective immediately. CFO Colette Kress is interim CEO. Nvidia gave no reason. According to AP, Huang had emailed staff days earlier apologizing for an affair, and Nvidia said it was checking whether he was “the target of a foreign intelligence effort”, with “no reason to believe” any secrets were compromised. The New York Times reported more claims that have not been confirmed. Nvidia’s stock fell 12.75% over the week, CNBC reports.
Why it matters: almost every AI product depends on Nvidia’s chips, so a sudden change at the top shakes the whole industry.
Nvidia’s announcement · AP via PBS · CNBC
Notion AI chats leaked through a vendor
Attackers broke into Fireworks AI, a company that runs AI models for other businesses. Notion says the stolen data included Notion AI chat transcripts from fewer than 1% of its users, even though Fireworks had agreed to keep no Notion data. Notion cut Fireworks off and tells affected users to change any password or key they typed into those chats.
Why it matters: what you type into an AI chat can end up stored by companies you have never heard of.
Grok Build apps were open to anyone
Security firm Claroc found 35,578 apps made with Grok Build that anyone could open without signing in. Some held hospital records, staff lists and plain-text passwords. Claroc says the publish step was set to “Anyone with the link” by default. The company behind Grok says apps stay private until their maker shares them.
Why it matters: if you build something with an AI app builder, check who can actually open it.
Claude Code saved passwords into public code
Security firm Apiiro found that clicking “Yes, and don’t ask again” in Anthropic’s Claude Code can save the approved command, password included, into a project file. In public GitHub projects they counted 2,539 leaked secrets in 1,994 repositories. According to The Hacker News, a bug report about this from August 2025 was closed by an automated bot as “not planned”.
Why it matters: AI coding tools can quietly write your secrets into files that get published.
United Airlines blocks AI agents
United said it will not let third-party AI agents book flights or manage reservations for you, on its own site or through travel agencies. Its reasons: handing your login to an AI tool can expose personal data, the tool might buy or change things you didn’t mean to, and federal rules need a person to finish check-in.
Why it matters: AI assistants that act for you will run into more closed doors like this one.
Nonprofits sue to stop frontier AI
On November 12 a group of nonprofits, including Public Citizen, sued OpenAI, Anthropic, Google, Meta and SpaceX’s AI division in San Francisco. They ask the court to stop the companies from building more advanced models, arguing it creates a public nuisance under California law. These are claims in a lawsuit, and there is no ruling yet.
Why it matters: courts are becoming one of the places where the limits of AI get decided.
Australia considers criminal charges for AI companies
Australia’s government said on November 10 it will consider a criminal offence for AI companies whose models cause harm when they didn’t take reasonable steps to prevent it, and an AI safety watchdog that can fine them. It follows a September incident that linked OpenAI to a hack of a Medicare statistics portal. Federal police say they have opened no investigation.
Why it matters: governments are starting to treat harm caused by AI as something a company could be prosecuted for.
US task force: go easy on companies hit by AI attacks
The White House AI task force recommended that agencies hold back on penalties for companies that report security incidents in good faith, and on enforcement against companies hit by AI-driven cyberattacks, for at least 12 months. These are recommendations, not a law.
Why it matters: the US is choosing to get companies to report AI attacks by easing penalties.
Microsoft 365 and Claude outages
From November 12 a global Microsoft 365 outage took down Outlook, Teams and Microsoft 365 Copilot. Microsoft said servers became unexpectedly unavailable and found no sign of an attack. It also broke Claude’s Microsoft 365 connector for about a day. Separately, on November 13 Claude had login failures and errors of its own, now resolved.
Why it matters: AI assistants built into work tools go down when the cloud underneath them does.