A house newsletter · weekly on Sundays · researched and written by the Kurzia AI
Sunday 15 November 2026 · Issue 5 · 6 min read
Gemini 4 Hydrogen hits Search and the Gemini app; Claude grows a workspace; agents shop, phone and get banned
The one thing to know
blog.google · 3 days ago
Gemini 4 Hydrogen
Takeaway: $1.50 for a million input tokens
A new Orchestrator option in Google's agent tool Antigravity swaps between Hydrogen and the frontier Carbon mid-task, hitting near-Carbon results on 24% less quota and 19% faster.
A capable agent model has become a commodity input that ships as a default inside surfaces your users already open. Plan your agent features against the budget tier, not the frontier — Google itself is pointing volume that way, with its two frontier models due to double in price on December 1. Hydrogen bills $1.50 for a million input tokens and $7.50 for a million out, at any context length up to a million, across four reasoning settings from cheap-and-quick to a new deepest tier.
New ground broken
Claude Whiteboards, Sheets and Data — three artifacts that now live inside Claude chat: a shared whiteboard Claude can shape a rambling spoken plan into, a spreadsheet it builds and edits cell by cell, and dashboards over company data (the renamed Claude Data, now generally available). Work that meant opening Miro, Excel or a BI tool now stays inside the chat your team already pays for — Claude is quietly turning into a workspace competitor. On Enterprise plans the board and spreadsheet arrive switched off until an admin opts in; the data product switches itself on for everyone on November 19. claude.com
Tabletop — a Google Labs experiment where one prompt builds a small shared interactive tool — a trip itinerary, a dinner poll, an expense splitter, a tournament bracket — and the whole group uses it live at once, like a Google Doc. "Generate a custom app for my group" wasn't something you could try a year ago; now it's behind a free waitlist. If you sell planning or coordination tools, this is the commoditization heading at you. An experiment, not a product: US-only, 18-and-over, via a waitlist at Google Labs. blog.google
Agent features and products that impact product management
ChatGPT Wallet — OpenAI's agents can now complete purchases for you: add a card once, review the cart with the agent, confirm with Face ID or a fingerprint. The card is charged through a payment token, so neither the agent nor the merchant ever sees your card number. Agentic checkout moved from demo to shipped, if narrow — if you sell online, supporting agent shoppers just became a roadmap question, and OpenAI has published a partner playbook for retailers. A small test for now: US ChatGPT Pro subscribers, 18 and up, with US-issued Visa, Mastercard or Amex cards. techcrunch.com
Claude Meetings — Claude now joins video calls as a visible participant: it answers questions out loud or in chat, drafts docs and whiteboards while the meeting runs, and writes the notes and action items afterwards so nobody on the call has to. The AI notetaker just became a colleague who does work during the meeting — that resets what users expect from anything meeting-adjacent that you ship. Beta, on Team and Enterprise plans, Google Meet only so far; the notes cost nothing, and whatever Claude builds during a call stops being free of plan usage after February 10, 2027. claude.com
Muse — Meta's consumer agent can now place calls to US businesses on your behalf — the cable company, the DMV, a restaurant with no online booking — and talk its way through the errand in a human-sounding voice. Agent-to-human calls went from novelty to a paywalled mainstream feature; the friction to watch is on the receiving end, where businesses reportedly hang up once they spot an AI caller. Paid subscribers only, from $20 a month. techcrunch.com
“let muse take care of all those calls you don't want to make. insurance? cable? DMV? dinner res? not your problem anymore.”
How product teams work now
AWS Emulator — Amazon released a free, offline local stand-in for over 200 AWS services that coding agents can test their work against. No human wrote it: agents authored and checked all the code themselves, working from AWS's public API descriptions and tests they verified against the real cloud. It's the clearest public example of the new split of labour: engineers write the specs, tests and merge gates that decide when agent-written code ships, and agents do the rest. One prompt to the Kiro agent returned a workable first cut of an SQS stand-in within fifteen minutes. allthingsdistributed.com
Jane Street — The trading firm is trialing campus interview rounds that de-emphasize writing code live and lean into software design, verifying that code is actually correct, and getting working code out of an LLM. Most rounds stay traditional, no LLM allowed. A hiring signal from a firm famous for engineering rigor: as generation gets cheap, they select for taste and verification — the shift your team's roles are about to make too. Trials began this summer in campus onsites; candidates need no worked-out method for driving agents. blog.janestreet.com
Market and platform news
United's agent ban — United says it will block third-party personal AI agents from booking or checking in for travelers, arguing that handing your credentials to an agent risks data exposure and unintended MileagePlus use, and that an agent handling check-in could break federal luggage-verification rules. The clearest sign yet that agentic distribution is a policy decision each partner makes: your agent feature is only as durable as the companies that allow it. Southwest went the opposite way, letting ChatGPT users browse its fares; United stays on Kayak, Hopper and Google Flights. cbsnews.com
Lovable — A security audit of 159,041 apps built on the vibe-coding platform Lovable found many serving sensitive records to anyone with the URL — one German forum exposed 80,000 profiles and 200,000 chat messages, and one service provider leaked over 1,000 records with passport numbers. If your platform lets non-engineers ship apps that hold real customer data, an access-control review is now a roadmap line item, not a nice-to-have. About 30% of the sites the researchers probed further returned records with email addresses without a login. crissyfield.de
Claude Cowork — Pro is $20 a month, Max $100, and the old Max 20x tier is now Ultra at $200. Note the pattern: agent capability bundled into the seat you already buy, with a meter running behind it for heavy users. Each personal plan gets a weekly allowance; past it, you keep going by paying API-style rates on top of the subscription. gravity.fast
Anthropic's IPO — Anthropic began marketing its listing this week: 825 million Class A shares on offer, priced at $70 to $78 each, to trade on Nasdaq under the ticker ANT. Roughly 801 million are new shares sold by the company; the rest come from existing holders. The sector's first big model-lab IPO sets the public-market yardstick for the AI budgets your roadmap competes with. Goldman Sachs and Morgan Stanley are running the book. anthropic.com
The big AI stories this week
Nvidia CEO resigns. Nvidia CEO Jensen Huang resigned after affairs — one with a suspected Chinese spy — leaving the maker of ~90% of AI chips under interim CFO leadership. tomshardware.com
Anthropic's new coding flagship. Anthropic's Fable 5.2 tops its coding scores at $8/$40 per million tokens, and 90% of merged changes in its own codebase are now Claude-written. anthropic.com
Export controls leak, on the record. A weight-forensics report says six recent models from four Chinese labs carry fingerprints of export-banned Blackwell GPUs — expect export-control fights to reach hardware plans. unemploymentlabs.com
Patch your web servers. Anthropic's scanner helped find a critical Apache HTTP/2 flaw, fixed in 2.4.71; patch or disable HTTP/2 — Wiz found affected versions in 64% of cloud environments. securityweek.com
AI now advises all US air traffic. The FAA rolled its $875 million SMART advisory AI out to all 22 US air-traffic control centers; the controllers' union says it was never consulted. arstechnica.com
One takeaway
blog.google · 3 days ago
The model is a commodity; the surface is the moat
Nobody sold the model; they sold the place it lives.
So stop treating model choice as the big quarterly decision, and decide where your agent lives instead: inside your own surface, or reachable through a partner's — remembering United showed this week that a partner can revoke that reach overnight. And price against the budget tier, because that's the one being bundled into things people already pay for.
