ANG⁺ #7 · Agent Wallets, +5 More
ChatGPT now has a wallet that lets agents pay for you. Plus Google's Gemini 4 Hydrogen, Claude joining your meetings, agents working hospital software, Hugging Face's routing switch, and Copilot models retiring December 14.
ANG⁺ #7 · Agent Wallets, +5 More
about 4 min
ANG⁺ = briefings sorted by AI and checked by the editor
Welcome to the seventh ANG⁺ Briefing. Briefing #6 (November 12) covered where money and work are moving. This edition is about where agents have started to act on their own. Between November 12 and 16, agents paid with real cards, joined video calls as participants, and clicked through hospital software. The same week, Google put a cheaper new default model into Search. Everything is built on first-party material: official company announcements and developer changelogs.
1. Google's Gemini 4 Hydrogen — a cheap default model in the API, the app and Search at once
On November 12 (U.S. time), Google made Gemini 4 Hydrogen (gemini-4-hydrogen) generally available. If Carbon and Argon from #6 are the models for digging deep, Hydrogen is the everyday one.
- Price:
$1.50per million input tokens and$7.50per million output tokens, flat across its full 1 million-token context. The Gemini API also has a free tier. - Size and thinking levels: 1 million input tokens and up to 1 million output tokens. Thinking levels are low, medium, high (default), plus a new
xhightier — four in all. - Google's numbers: On the Artificial Analysis Intelligence Index it scores 51 at its default level, above Argon (50), and 54 at xhigh. Against the previous Gemini 3.8 Flash, Terminal-Bench 4.1 goes from 18.2% to 57%, and tasks finish 2.5 times faster.
- Where: the Gemini API, AI Studio, Gemini for business and the Gemini app (free users get it through auto model selection). In Google Search it reached all users worldwide, in every language AI Mode supports. GitHub Copilot made it generally available on November 13 for Pro, Pro+, Max, Business and Enterprise.
Our take: Carbon from #6 costs$2input and$10output at its introductory price until November 30, then$4and$20. Hydrogen starts at$1.50and$7.50. If you are on Gemini 3.8 Flash today, run the same job once on Hydrogen before December and compare cost and results.
Source: Introducing Gemini 4 Hydrogen for Agentic Workflows (Google) · Google Search upgrade: Gemini 4 Hydrogen features (Google) · Gemini API changelog, November 12, 2026 (Google) · Gemini 4 Hydrogen in GitHub Copilot (GitHub)
2. OpenAI's ChatGPT Wallet — a wallet that lets agents pay for you
On November 13, OpenAI introduced ChatGPT Wallet, which lets an agent pay on a user's behalf. It was built with American Express, Chase, Mastercard and Visa.
- How one purchase works: You add a card once. Before every purchase you review the items, the total and the selected card, and approve each one with a security check such as Face ID or a fingerprint.
- No card number handed over: After approval, the payment network issues a payment token that stands in for the real card details. Neither the business nor the agent gets the actual card number, and the card's existing rewards and dispute protections stay in place.
- Who can use it first: It starts in dots, for ChatGPT Pro subscribers aged 18 and older in the U.S. with eligible U.S.-issued cards.
- A playbook for businesses: The "Commerce Partner Playbook" released alongside it covers publishing product information, keeping agents from being blocked (recognizing signed requests with Web Bot Auth), and working with payment and risk providers. OpenAI named fraud-prevention partners (Forter, Riskified, Sardine, Signifyd) and payment providers (Adyen, Checkout.com, Stripe).
Our take: If "Businesses in ChatGPT" from #6 was the front door that helps customers find a store, this wallet is the checkout counter. If you run an online store, check this week that your firewall or CDN is not blocking agents' signed requests.
Source: Introducing ChatGPT Wallet (OpenAI)
3. Anthropic's Claude Meetings beta — a "visible participant" in your calls
On November 12, Anthropic began rolling out Claude Meetings in beta, which puts Claude directly into video calls. It does more than take notes — it starts work during the meeting.
- What it does: Claude joins as a visible participant, follows the conversation, answers questions, builds docs or whiteboards with teammates, and leaves notes and action items.
- Who and where: It is rolling out in beta on Team and Enterprise plans. For now it works only in Google Meet, with other platforms "coming soon."
- Price: Notes and transcripts are free during the beta. Answers in calls are free too, unless an answer uses a connected tool, which counts toward plan usage. Docs, whiteboards and other artifacts are free through February 10, 2027.
- Where it is not available: Organizations using zero data retention (ZDR), a HIPAA-ready configuration, or customer-managed encryption keys (CMEK).
Our take: The meeting assistant is moving from "the one who takes notes" to "the one who works alongside you." What a team needs to settle before adopting it is a rule, not a setting — write down first how participants will be told that an AI is in the meeting.
Source: Introducing Claude Meetings, now in beta (Anthropic)
4. OpenAI's "ChatGPT for Healthcare" — agents that work hospital software
On November 16, OpenAI added computer-use agents to ChatGPT for Healthcare. It comes one week after Google's "Gemini for healthcare," which we covered in #6.
- What changes: Agents read screens, click, type and move between the programs a hospital already uses — processing referrals or moving documents between systems. Organizations decide which applications agents can reach and which steps need human review, with access controls and audit logs.
- Who is using it: OpenAI highlighted work with Abridge, Ambience, Cigna Healthcare and CVS Health.
- Answer quality: In over 1300 blind reviews by physicians, the updated version was rated higher on all eight dimensions evaluated.
- Plugins: More than 20 new healthcare plugins, including UpToDate Expert AI, Oracle HDI and Trizetto. Verified U.S. clinicians can use ChatGPT for Clinicians for free, and the computer-use capability is also in the OpenAI API, ChatGPT Work and Codex.
Our take: Both companies walked into hospitals in the same month. What will separate them is less the model score than how finely you can set "where does a human stop and check." If you are evaluating either one, look at that settings screen first.
Source: More ways to move care forward with ChatGPT for Healthcare (OpenAI)
5. Hugging Face routes language-model requests through OpenRouter — your own API key no longer works
On November 16, Hugging Face switched Inference Providers' language-model requests from its own router to OpenRouter, which spreads requests across 80+ providers.
- What gets better: Before, if one provider failed, you got its error. Now the request is retried on another provider. Hugging Face said capacity errors fell 46%.
- What breaks: For language models, you can no longer use your own provider API key. Keys for image and video providers keep working. You can still pick a provider with a
model:providersuffix. - Price: Unchanged. PRO users get
$2of credits every month, and Team and Enterprise organizations get$2per seat. - Numbers shared alongside: On OpenRouter, open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026. Average prices per million tokens:
$0.28input and$1.62output for open-weight models, versus$1.27and$6.36for closed ones.
Our take: If you have code that calls language models through Hugging Face with another company's key, it needs fixing today. On the numbers, the big one is that open models already serve more than 70% of tokens. At around a quarter of the price, "try an open model first, step up if it falls short" is becoming a sensible default.
Source: Hugging Face and OpenRouter join forces to make open-weight models easy to run (Hugging Face)
6. New this week · dates for your calendar
Small changes from November 12–16, plus dates within a month when prices or code change.
- OpenAI brings outside assessors inside (November 13): It signed "embedded assessment" agreements with four organizations — Apollo Research, METR, Transluce and Goodfire. Assessors get access to model checkpoints, source code, and training and evaluation data, among other things; METR will assess OpenAI's largest reinforcement-learning run to date.
- Self-hosted Claude Managed Agents (November 12): Not just sandboxes — the whole service and session history can now run on your own infrastructure. Access is requested through your Anthropic account team.
- Claude for Manufacturing (November 12): One manufacturing plugin with 17 skills and 4 connectors (Autodesk Fusion, Infor, Microsoft 365, Oracle NetSuite), for Team and Enterprise plans. The technical webinar is November 17 at 11:00 a.m. ET.
- DeepSeek puts its app features in the API (November 13):
deepseek-flash(V4.1-Flash) now supports web search, deep-thinking search, file input, URL reading and safety checks, and Search, Fetch, OCR, ASR and TTS can be called on their own. Public-facing apps must turn safety checks on. - llama.cpp and ggml
v1.0.0(November 13): The base libraries of local AI reached 1.0. Breaking changes will now bump the major version; they support more than 150 architecture variants and ship through Homebrew, conda-forge, Winget and more. - Google DeepMind opens its weather-AI datasets (November 13): SatWeave (satellite imagery, 17 channels, updated hourly) and PARDIG (precipitation maps at 15-minute intervals, about 10 km resolution, records from 1999 to 2026) are on Google Cloud Storage. PARDIG improves hourly detection of light-to-moderate precipitation over NASA's IMERG-Final by more than 30%.
- November 30: Gemini 4 Carbon's introductory price ends (then
$4input,$20output) · Claude Sonnet 4.5 retires from the API · fast mode is removed for Claude Opus 4.8 (afterwards,speed: "fast"requests return an error). - December 1:
gemini-2.5-proandgemini-2.5-flashauto-switch (#4). December 3:gemini-omni-1.1-flash→gemini-omni-1.2-flashauto-switch (#5). - December 14: Six GitHub Copilot models retire — Claude Fable 5 → 5.2, Claude Haiku 4.5 → 5.5, Gemini 3.8 Flash → 4 Hydrogen, Gemini 4 Argon → 4 Carbon, Grok 4.6 → 4.7, Kimi K3 → replacement to be announced. Haiku 4.5 stays for individual subscribers on annual plans.
Our take: What this week's small items share is "where, and under whose eyes, do agents run" — on your own servers (Managed Agents), under outside assessors (OpenAI), with safety checks on (DeepSeek). On the calendar, just two dates need to go in first: November 30 and December 14.
Source: Our first phase of embedded assessments (OpenAI) · Claude release notes, November 12, 2026 (Anthropic) · Claude for Manufacturing: AI plugin for plant teams (Anthropic) · The full DeepSeek App experience, now via API (DeepSeek) · llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem (Hugging Face) · PARDIG & SatWeave: Foundational Datasets for Earth Science (Google DeepMind) · Upcoming deprecation of selected GitHub Copilot models in December (GitHub)