Quick answer
ChatGPT Wallet is OpenAI’s new payment feature that lets ChatGPT’s “dots” agents buy things for you without seeing your real card details. You review every purchase and approve it with Face ID or a fingerprint, and the shop receives a one-off payment token instead of your card number. It launched on 13 November 2026 for ChatGPT Pro subscribers in the US only. The same week, a new Gemini 4 Krypton test version was reported (unconfirmed), Claude Code got custom themes and a big Projects update, Anthropic published advice on which Claude model to use, Google reported maths results from its Gemini agents, and a GLM-5.5 “leak” claimed Opus-level scores that may be fake.
Key takeaways
- ChatGPT Wallet is real and official, but US-only and Pro-only for now. Every purchase needs your approval.
- Gemini’s Google Wallet link is read-only. Google says it gives spending insights and cannot move your money, which is different from how it was described in the video.
- Gemini 4 Krypton is still a rumour. The demos look strong, but Google has not announced it.
- Claude Code Projects now has a more proactive coordinator, /btw, /fork, thread reminders and read-only share links.
- Anthropic says start with Opus 5.5 at medium effort, and raise the effort before paying for a bigger model.
- Treat the GLM-5.5 and GPT-6.2 Astra material as leaks until the companies publish something.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Let an AI Agent Make Purchases for You Safely — spending limits, approval steps, virtual cards and receipts before you hand any agent a card.
- How to Share an AI Chat Link Without Leaking Private Information — a five-minute check to run before you send anyone a link to a ChatGPT, Claude or Gemini conversation.
In this guide
- Watch: the source video
- What is ChatGPT Wallet and how does it work?
- Can Gemini buy things with Google Wallet too?
- What is new with Gemini 4 Krypton?
- What changed in Claude Code this week?
- What do the GPT-6.2 Astra previews show?
- Which Claude model does Anthropic recommend now?
- What did Google’s Gemini agents do in maths?
- What else happened in AI this week?
- What is confirmed and what is only claimed?
- What should you do about this week’s news?
- Frequently asked questions
- The bottom line
- Related posts
This week’s AI news had a theme running through it: agents are being handed more responsibility. They can now pay for things, coordinate several threads of work at once and even help prove maths problems. That makes it a good week to slow down, check what is actually confirmed and set some sensible rules for your own tools.
This round-up is based on the 15 November news video from WorldofAI, with each major claim checked against official announcements where we could find them.
Watch: the source video
What this video covers: WorldofAI’s weekend AI news round-up: a new Gemini 4 Krypton test version and its demos (a shooter game clone, a 3D rocket engine, SVG artwork), Claude Code custom themes and Projects updates, early GPT-6.2 Astra outputs, Google’s Gemini maths results, OpenAI’s robotics team changes, ChatGPT Wallet, Anthropic’s Claude model guide, Dario Amodei on the IPO, the GLM-5.5 leak, Gemini Notebook and Google Wallet, Antigravity workflows, DeepSeek API changes, Muse Spark 1.4, a Qwen 4 Max upgrade, Perplexity Teammates, SpaceX’s Starship catch and two Fable 5.2 animations made entirely in code.
Video by WorldofAI (@intheworldofai). The video was sponsored by Mammouth, an AI platform that bundles several models from $10 a month. We have no link with the sponsor.
What is ChatGPT Wallet and how does it work?
ChatGPT Wallet lets an AI agent pay on your behalf while keeping your card details hidden from both the agent and the shop. OpenAI announced it on 13 November 2026, built with American Express, Mastercard and Visa.
Here is how a purchase works, according to OpenAI and its help page:
- You ask your dot (ChatGPT’s agent) to find something, such as a gift or a hotel.
- At checkout, it shows you the items, the total and the card it will use.
- You approve with a security check such as Face ID or a fingerprint. Every purchase needs this step.
- The card network issues a payment token, a stand-in card number, so the business never sees your real card details.
- You keep your card’s usual rewards and dispute protections for eligible purchases.
| Detail | What OpenAI says |
|---|---|
| Who can use it | ChatGPT Pro subscribers in the US, aged 18 or over, with a US billing address |
| Cards | Eligible US-issued American Express, Chase, Mastercard and Visa cards |
| Where it works | In conversations with your dot only, not in normal chats or standalone Codex threads |
| Approval | A biometric check on every purchase |
| Coming later | Pre-approved purchases within limits you set, such as spending caps, categories or time periods |
In the video’s demo, an agent finds a holiday gift within a budget, suggests a cheaper option and asks for approval before paying. The creator pictures the same idea for groceries: the agent builds the list and you approve the payment.
⚠️ Warning: ChatGPT Wallet is not available in the UK yet. If someone offers you “ChatGPT Wallet access” outside the official app, treat it as a scam.
💡 Tip: Before you let any agent spend money, decide your rules first. Our guide How to Let an AI Agent Make Purchases for You Safely walks you through limits, approvals and receipts.
Can Gemini buy things with Google Wallet too?
Not in the way the video suggests. The creator described Google’s update as Gemini users connecting Google Wallet so Gemini can make purchases for them, like ChatGPT Wallet.
Google’s own announcement on 13 November says something narrower. Gemini app users in the US can connect Google Wallet to get personal finance insights, such as spending charts, forgotten subscriptions and holiday budgets. Google says that when connected, Gemini can only read your account information; it cannot move, manage or invest your money.
Google does have a separate agentic checkout feature with Google Pay for some shopping in the US, so the two ideas are related. But the Wallet connection announced this week is about understanding your spending, not handing over the card.
What is new with Gemini 4 Krypton?
🔎 Unconfirmed / rumour: Google has not announced Gemini 4 Krypton. Everything in this section comes from testers’ posts and the creator’s commentary, not from Google.
Gemini 4 Krypton is the rumoured next model in Google’s Gemini 4 family, after Argon, Carbon and the recently released Hydrogen. We covered the first Krypton reports in Gemini Krypton Leak: Claude Fable 5.2, a Hidden Fable 6 and This Week’s AI News. This week the creator says a newer test version was spotted on 13 November, reportedly being tried out on the Arena comparison site and through the API.
The demos shown in the video include:
- a playable shooter game clone with loadouts, a map and a health bar, made by a friend of the creator;
- an interactive 3D rocket engine with exploded, X-ray and thermal views, built in a single HTML file;
- a page-turning book animation in JavaScript, with a few visual glitches;
- SVG drawings of a game-style city scene and two games controllers.
The creator is balanced about it. He says it feels closer to Claude Sonnet 5.5 than Opus 5.5 for web design work, though it may beat Opus in other areas, and he does not think it matches Fable 5.2 overall despite posts claiming it is “Fable 5.2 level but cheaper”. Some testers say a release could come within a week or so. That is speculation.
If you want something you can use today, Google’s fast model Gemini 4 Hydrogen is already out.
What changed in Claude Code this week?
Two things: custom themes for the desktop app, and a larger update to Projects that makes working across several parallel threads easier.
Custom themes
You can now restyle Claude Code in the desktop app, including colours and animated backgrounds. You do not have to write the theme yourself; you can describe what you want, add reference images and ask Claude to build it, or start from Anthropic’s example themes plugin. The creator says it works for local sessions in the latest desktop app for now. It is a small change, but a pleasant one.
Projects: a more helpful coordinator
In Claude Code Projects, one coordinator hands work to several threads running at once. Anthropic’s update makes that coordinator more proactive: it answers questions in the project chat and reports what each thread is doing. Other additions:
- /btw opens a separate thread for an unrelated question, so your current thread stays on topic.
- /fork starts a new thread from the current one, so you can try a different approach without losing your progress.
- Thread reminders park a thread until a set time, then bring it back into the project chat.
- Read-only share links for individual threads, which anyone signed in to Claude can open. Threads that had access to your computer cannot be shared.
- A tidier threads panel with recent and pinned items, plus thumbnails of the files each thread made.
Projects is in public beta for Claude Pro and Max users, with a waitlist for those without access.
💡 Tip: Sharing a thread is handy for showing a colleague how you solved something, but a shared chat can contain more than you remember. Run through How to Share an AI Chat Link Without Leaking Private Information before you send one.
Try this today: open the AI tool you use most and check two settings: what it can spend or send on your behalf, and which of your chats are currently shared. Both take under five minutes.
What do the GPT-6.2 Astra previews show?
🔎 Unconfirmed / rumour: These are early-access outputs posted by testers on X. OpenAI has not released GPT-6.2 Astra or published results for it.
The previews suggest a model that is very good at detailed 3D scenes and visual design from a single prompt. The video shows a tester’s “detail study” scene with weathered metal, wet road reflections and fine hardware, and a side-by-side of three games controllers made by GPT-6.2 Astra, Gemini 4 Hydrogen and Gemini 4 Carbon.
The creator says a full video on access is coming. For our earlier look at the model’s leaked tests, see GPT-6.2 Astra: Early SVG Tests, OpenAI’s GPT-ADA Models and This Week’s AI News.
Which Claude model does Anthropic recommend now?
Anthropic’s advice is to start most work on Claude Opus 5.5 at its default “medium” effort, then move up or down only when you need to. The guide, published on the Claude developer blog on 9 November, suggests:
| Model | Use it for |
|---|---|
| Haiku 5.5 | Simple, high-volume jobs such as sorting, tagging and pulling out data, and as a helper for a bigger model |
| Sonnet 5.5 | Well-defined everyday work, such as a bug fix or a feature built to a clear spec, and fast chat |
| Opus 5.5 | Your default for most coding and knowledge work |
| Fable 5.2 | The hardest projects, long unattended runs, or when Opus at high effort still falls short |
The most useful detail is a chart shown in the video. Across 52 coding tasks in Claude Code, Opus 5.5 solved 84.6% at about $0.65 per solved task, while Fable 5.2 solved 82.7% at about $1.83, roughly 2.8 times the cost. For everyday work, the most expensive model is not automatically the best one.
Anthropic also suggests changing the effort level before changing the model. We cover that in Low, High or Max: How to Choose an AI Reasoning Effort Level, and our plain-English version of the model choice is in Which Claude Model Should You Use? Haiku, Sonnet, Opus or Fable.
What did Google’s Gemini agents do in maths?
Google reports that its Gemini 4 models, working as teams of AI agents alongside human mathematicians, produced machine-checked proofs and formal versions of several long-standing maths problems. The examples it gives include the first formal, computer-checked version of the Birkhoff conjecture (open since 1927) and of the 3D Bochner–Riesz conjecture, where the system also found and fixed a gap in a recent proof.
The creator also quotes very large figures for the amount of proof code and the number of agent runs involved. Some independent write-ups point out that not all of the formal proofs have been published for others to check yet, so treat this as Google’s report for now. It follows similar claims from OpenAI a few weeks ago.
What else happened in AI this week?
| Story | What was said | Status |
|---|---|---|
| OpenAI robotics team | The standalone robotics team has been split up, with researchers moved to the main research group and hardware staff to the applied group. OpenAI says general models will increasingly run robots. | Reported by Fortune (13 Nov) |
| Anthropic IPO | Dario Amodei says the main reason to go public is to fund computing power, and he is not fussed about the exact timing. | Interview clip; see our Anthropic IPO explainer |
| GLM-5.5 | A leaked internal chat and chart claim Z.ai’s next model beats Claude Opus 5.5. | Leak; the creator says it may be fake |
| Gemini Notebook | A comic-style artifact option was spotted in development. | Spotted, not launched |
| Google Antigravity | A new /run-workflow command turns a task into a Python script that runs several agents in parallel. You can review it first, save it and reuse it. | Feature update |
| DeepSeek API | App features such as web search, files, links and voice are coming to the API, plus separate OCR, speech recognition and text-to-speech APIs. | Announced |
| DeepSeek subscriptions | Code showing a usage-limit notice suggests paid plans could follow. | Speculation |
| Muse Spark 1.4 | Meta’s next model was reportedly spotted in internal code. | Report |
| Qwen 4 Max Preview | The creator’s own tests suggest a quiet upgrade. See our Qwen 4 preview first look. | Creator’s tests |
| Perplexity Teammates | Perplexity’s proactive agent feature comes to the Mac app for Max subscribers. | Feature update |
| SpaceX Starship | SpaceX reportedly caught the Starship upper stage at the launch tower for the first time, and Elon Musk said Grok helped with the work. | Reported; the Grok role is Musk’s claim |
| Fable 5.2 animations | A PC-building animation (about 7 hours 41 minutes and $138 in usage, according to its creator) and an anime-style short, both made with code. | Community demos |
What is confirmed and what is only claimed?
| Claim | Status | How we checked |
|---|---|---|
| ChatGPT Wallet launched for US Pro users | Confirmed | OpenAI announcement and help page |
| Gemini can make purchases through Google Wallet | Not as described | Google says the Wallet connection is read-only finance insights |
| Claude Code Projects update (/btw, /fork, reminders, sharing) | Confirmed | Anthropic’s announcement, 13 Nov |
| Anthropic recommends Opus 5.5 at medium as the default | Confirmed | Claude developer blog, 9 Nov |
| OpenAI disbanded its standalone robotics team | Reported | Fortune exclusive |
| Gemini 4 Krypton nearly matches Fable 5.2 | Unconfirmed | Tester demos only |
| GPT-6.2 Astra early outputs | Unconfirmed | Tester posts only |
| GLM-5.5 beats Opus 5.5 | Unconfirmed, possibly fake | Leaked chart |
| DeepSeek paid subscriptions coming | Speculation | Code spotted on the website |
For leaks like GLM-5.5, our checklist How to Spot a Fake AI Model Leak Before You Download Anything will help you avoid fake downloads and fake “early access” offers.
What should you do about this week’s news?
- Set spending rules before you need them. Agent payments will reach the UK eventually. Decide now what an agent may buy, the maximum amount and who approves it. Our guide How to Let an AI Agent Make Purchases for You Safely has a template.
- Check your shared chats. Most AI tools now let you share conversations. Review which links exist and what is in them. Use our five-minute check.
- Don’t switch models on a rumour. Wait for Krypton, Astra or GLM-5.5 to be released and test them on your own work.
- Match the model to the job. If you use Claude, try Opus 5.5 at medium effort before reaching for Fable 5.2.
- Review what your agents can reach. Payments, email and files deserve the most care. Our 30-minute AI access audit is a good starting point.
Your next step: write down three rules for any AI agent that handles money in your business: what it can buy, the most it can spend, and who must approve it. Keep them where your team can see them.
Frequently asked questions
What is ChatGPT Wallet?
It is a payment feature that lets ChatGPT’s dots agents buy things for you. You approve each purchase with a biometric check, and the business receives a payment token instead of your real card details.
Is ChatGPT Wallet available in the UK?
No. At launch it is only for ChatGPT Pro subscribers in the US with eligible US-issued cards and a US billing address.
Can Gemini spend money from my Google Wallet?
Not through the connection Google announced on 13 November. Google says it is read-only and cannot move, manage or invest your money.
Has Gemini 4 Krypton been released?
No. It has only been reported through testers and leaks. Google has not announced it.
Which Claude model should I use for coding?
Anthropic recommends starting with Opus 5.5 at medium effort, using Sonnet 5.5 for well-defined tasks and Fable 5.2 for the hardest projects.
Is the GLM-5.5 leak real?
Nobody outside Z.ai knows yet. Even the creator warns it may be fake, so wait for an official release and independent tests.
Was the video sponsored?
Yes, by Mammouth, an AI platform offering several models from $10 a month.
The bottom line
ChatGPT Wallet is the story to watch, not because you can use it in the UK today, but because it shows where AI tools are going: from suggesting things to doing them, including spending your money. OpenAI has built sensible guard rails, such as approval on every purchase and hidden card details. Your job is to add your own. Treat the Krypton, Astra and GLM-5.5 chatter as interesting but unproven, and spend your time on the confirmed tools you already have.
Related posts
- Gemini Krypton Leak: Claude Fable 5.2, a Hidden Fable 6 and This Week’s AI News
- Which Claude Model Should You Use? Haiku, Sonnet, Opus or Fable
- If a Staff Laptop Is Hacked, What Can Your AI Tools Do? A 30-Minute Access Audit
