Agents go shopping, sellers push back

The week in AI for business leaders. Agents reached the checkout: OpenAI's Business Hub and its new ChatGPT Wallet on one side, United, eBay and Amazon deciding who gets in on the other. Then the small, cheap decision models from Amazon, Google, Meta and Perplexity that are taking over routine sorting work. And a run of leaks, from Claude Code settings files and Grok Build share links to stolen AI keys at Railway and the Fireworks breach that reached Notion AI chats. Gritt Scott, who writes Skill Issue, and North Wayne, who writes First Opinion, join us. Plus the dates to watch.This week's stories00:31 OpenAI's Business Hub lets businesses claim their ChatGPT listing; eBay appears to block outside shopping agents and Amazon already does; United stops unauthorized AI agents from booking, so a person must finish the purchase; Muse now calls US businesses. After Friday's show, OpenAI launched ChatGPT Wallet: its dots agents can complete purchases for US ChatGPT Pro users, with a biometric approval for each one and a payment token instead of the real card number, plus a Commerce Partner Playbook for stores. The Verge, citing Reuters, reports an internal Amazon plan to index products sold anywhere; Amazon calls it a brainstorming document. Gritt Scott, who writes Skill Issue, on setting your rule for agents on purpose, talking to your payments and fraud providers, keeping your product listing accurate, and a one line policy for agents spending company money. (OpenAI (Wallet), TechCrunch (Wallet), The Verge (Wallet), OpenAI (Businesses), United, The Verge (eBay), TechCrunch (Muse calls), The Verge (fake shops study), The Verge (Amazon))04:44 Amazon engineers swapped Claude Sonnet for a small decision model and cut ticket labeling from about $22,200 to about $500 per million tickets. Google launched the Gemini Decisions API at $0.10 per million input tokens, Meta released Muse Ember, Google made Gemini 4 Hydrogen its fast everyday model at $1.50 per million input tokens, and Perplexity released Decider v1.3 as open weights (its test results are its own). Ramp's October index: the average price per million tokens fell 21% in a month while AI spend rose 1.65%. North Wayne, who writes First Opinion, on finding the sorting jobs, asking vendors which model sits underneath, using confidence scores to send unsure cases to a person, and testing before switching. (Strands Agents (Amazon), Google (Decisions API), Meta (Muse Ember), Google (Gemini 4 Hydrogen), Perplexity (Decider v1.3), Ramp AI Index, Anthropic (Fable 5.2))07:58 Apiiro found over 100,000 Claude Code settings files in public GitHub repos, about 1,100 from 63 organizations holding passwords or keys. Claroc found more than 43,000 Grok Build app links, over 35,000 open to anyone. Railway says stolen AI provider keys ran up about $740,000 in charges and it will cover what providers won't refund. Notion's latest Fireworks update says it revoked Notion keys found in affected chats and reported other companies' keys to their issuers; its investigation is ongoing. Set spending limits on AI keys, switch AI built apps to invited people only, check for .claude folders in public code, ask vendors which AI companies see your data, and keep passwords out of AI chats. (Ars Technica (Claude Code), Ars Technica (Grok Build), Railway, Notion, BleepingComputer)Dates to watchNovember 19: Claude Data turns on by default for Claude Enterprise. (Claude)November 20: Conductor starts deleting cloud data. Export now if your team used it. (Conductor)End of November: Gemini 4 Carbon's introductory price doubles. (Google)Nvidia after Jensen Huang's resignation: the board says customer commitments and the roadmap continue without interruption. (NVIDIA)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Practical AI news for people running businesses. Today, Google puts its new everyday Gemini model in front of almost everyone, a breach at an AI supplier most people never heard of reaches Notion users, and a 145 billion parameter model claims it runs on your laptop. Nix Nullty, who writes Overhyped, joins us on the laptop claims. Plus three quick hits, and what each one means for you.Today's stories00:21 Google releases Gemini 4 Hydrogen, its fast everyday model, in the Gemini app (free users via auto model selection), Google Search worldwide, Gemini for business, Gemini in Sheets and the Gemini API, replacing Gemini 3.8 Flash. Google says it is 2.5x faster than 3.8 Flash with more than 5x fewer fabricated claims; Artificial Analysis puts it at 51 on its Intelligence Index, just above Gemini 4 Argon. $1.50 per million input tokens and $7.50 per million output. Test your own automations on real work and compare the bill before switching, and watch search traffic as Google links to a wider range of sites. (Google, Google Search, Ars Technica, TechCrunch)02:04 An attacker reached a Fireworks AI development system through its cloud provider. On November 8 Fireworks confirmed the stolen data included Notion AI chat transcripts used for Notion's Insights feature. Notion says fewer than 1% of workspaces are affected and most transcripts carry no user or workspace labels; it has disabled Insights, stopped using Fireworks, and tells affected customers to treat the chats as exposed and rotate any credentials pasted in. Ask your vendors which AI providers see your data and which features store chats, and keep passwords and keys out of AI chats. (Notion, TechCrunch, BleepingComputer, VentureBeat)03:38 Edge0 open-sources Edge0-145B, a 145 billion parameter model with about 10 billion active per step, which it says runs on a compatible laptop at 29GB peak memory and 80+ tokens per second, and keeps up with named cloud models on coding and reasoning. The weights and code are public, but the benchmarks and speed tests are the company's own and no outside tests are out yet. Nix Nullty, who writes Overhyped, on what machine it really needs and how to test it. (Edge0, Hugging Face, LavX News)Quick hitsJensen Huang resigns as Nvidia's CEO, president and board member; CFO Colette Kress is interim CEO and Harvey Jones becomes the board's first chair. The board says customer commitments and the product roadmap continue without interruption. (NVIDIA, SEC filing, TechCrunch)A Podqi study shared with The Verge found risky or fake retail sites in 15% of Perplexity, 8% of ChatGPT and 4% of Gemini shopping answers across 151 brands, with smaller brands hit about twice as often. (The Verge)Meta's Muse assistant can now place phone calls to US businesses for paying subscribers, for calls such as cable companies, insurers and restaurant bookings. (TechCrunch)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Practical AI news for people running businesses. Today, United Airlines shuts the door on AI agents booking flights, new payment data shows what businesses actually spend on AI, and Google launches a cheap tool for quick yes or no decisions. Rex Factor, who writes Proof of Work, joins us on the spending story. Plus three quick hits, and what each one means for you.Today's stories00:20 United Airlines is restricting unauthorized third-party AI agents and assistants from booking and managing trips, on its own website and app and through travel agencies and aggregators. Agents can still search and show flights, but a person must complete the purchase. United cites account security, unintended purchases and FAA and DOT rules; CEO Scott Kirby told the WSJ he doesn't want "another middleman". If you sell online, decide your own rule for agents before they arrive at checkout. (United, United FAQ, The Verge, CNBC)02:07 Ramp's October AI Index, from card and bill payments at 85,000+ US businesses: AI spend up 1.65% to a new high; the median firm pays $16.32 per employee per month, the top 1% a median $10.2K; 1% of businesses drive 80% of OpenAI and Anthropic revenue; engineering is 72% of classified spend, while customer support rose from 0.8% in June to 13.3% in October; average price per million tokens fell 21% in a month. Rex Factor, who writes Proof of Work, on how far to trust the data and how to price a support chatbot. (Ramp, Ramp AI Index, The Decoder)04:17 Google launches the Gemini Decisions API with Gemini Decision 1: up to 1,000 yes/no, multiple choice or rubric questions against one input (text, images, audio, video, PDFs), with a confidence score per answer, in 150 to 250 ms, at $0.10 per million input tokens and nothing for output. Good for small back-office checks; send low-confidence answers to a person, and ask vendors whether the saving reaches your bill. (Google, Google AI Studio docs)Quick hitsOpenAI retires ChatGPT's Advanced and Standard Voice on December 11 and moves everyone to Live Voice; live video and screen sharing stay on Advanced Voice until Live supports them. (OpenAI Help Center, Android Authority)Databricks unveils an always-on Genie agent that works in the background on longer jobs and asks people for decisions; Genie Ontology is generally available, and the always-on mode goes to preview and beta first. (Databricks, TechCrunch)Meta releases Muse Ember, its smallest and cheapest model, for repetitive agent work like extraction and routing, at $0.10 per million input tokens and $0.50 per million output. (Meta, Meta pricing)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Practical AI news for people running businesses. Today, ChatGPT lets businesses claim their own listing, Anthropic turns Claude into an office suite, and thousands of apps built with Grok Build turn out to be open to anyone. Gritt Scott, who writes Skill Issue, joins us on the Anthropic story. Plus three quick hits, and what each one means for you.Today's stories00:20 OpenAI launches Businesses in ChatGPT in beta. Through Business Hub, a verified business can update its details, add product feeds with prices and availability, and see how often ChatGPT mentions it and sends clicks. Eligible US online stores and online service businesses can also launch a Business Agent that answers customer questions, and sponsor it as a ChatGPT ad. Claim the listing now; test the agent hard before it speaks for you. (OpenAI, The Verge)02:18 Anthropic launches Claude Sheets and Claude Whiteboards in beta on all paid plans, makes Claude Data generally available with certified queries, and, per TechCrunch and The Verge, puts Claude Meetings in beta on Team and Enterprise plans for Google Meet. Gritt Scott, who writes Skill Issue, on which tools to keep, rules for a bot in your meetings, and the admin defaults (Claude Data turns on by default for Enterprise on November 19). (Anthropic, TechCrunch, The Verge)04:45 Security firm Claroc found more than 43,000 Grok Build app links through public DNS records, and over 35,000 opened for anyone, exposing business plans, worker medical exams and sales prospect lists, including 24 apps from the vendor's own staff. The vendor says apps are private unless shared. Lovable, Bolt and Replit use the same kind of per-app address. Check who has shared AI-built apps and switch them to invited people only. (Ars Technica)Quick hitsGoogle's Gemini 4 Carbon is available to paying developers and businesses at an introductory $2 per million input tokens and $10 per million output, doubling at the end of November; it uses more tokens than Argon, so compare the full bill. (Google, The Verge)Railway says attackers who broke in last month stole customers' AI provider keys and ran up about $740,000 in charges; it will cover what providers won't refund. Set spending limits on any AI keys you store on a platform. (Railway)eBay appears to be blocking third-party AI shopping agents, including Meta's Muse, from browsing listings; Amazon already blocks them entirely. (The Verge)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Practical AI news for people running businesses. Today, Anthropic's new top model arrives at a lower price, Amazon cuts an AI bill by 98% with a much smaller model, and Claude Code settings files leak passwords onto GitHub. North Wayne, who writes First Opinion, joins us on the Amazon story. Plus four quick hits, and what each one means for you.Today's stories00:20 Anthropic releases Claude Fable 5.2 at $8 per million input tokens and $40 per million output, up to 30% faster than 5.1, with Mythos 5.2 limited to vetted organizations. Anthropic's own advice is to start most work on Opus 5.5, its own table has Gemini 4 Carbon slightly ahead on business workflows, and prompts are kept 30 days for safety monitoring unless you have zero data retention. (Anthropic, The Verge)02:15 Two Amazon engineers replaced Claude Sonnet 4.6 with Strands Decider, a small decision model, for labeling tickets in a huge internal system: about $500 instead of $22,200 per million tickets, and calls under one second. North Wayne, who writes First Opinion, on running old and new side by side, when this approach won't work, and asking vendors what model sits under simple sorting jobs. (Strands Agents)04:43 Apiiro found over 100,000 Claude Code settings files in public GitHub repos, about 1,100 from 63 organizations holding passwords, API keys or private keys saved by "don't ask again" approvals. If anyone builds tools for you with Claude Code, have them check for a .claude folder and change any exposed passwords. (Ars Technica)Quick hitsChatGPT Enterprise is now sold through AWS Marketplace, and eligible spend can count toward AWS commitments. (OpenAI)Cisco AI Defense and IBM Concert are adding Anthropic's Mythos to their security tools; Cisco says at no extra cost under existing licenses. (The Register)Apple sells $49.99 Intelligence Usage Packs in Logic Pro, with more apps including Pages, Keynote and Numbers expected this month. (The Verge)OpenAI says Costs API exports published November 6 to 9 had incorrect cost data; it reports the issue resolved. Re-pull any AI spend reports from those days. (OpenAI Status)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Welcome to the first daily broadcast of AI Squawk: practical AI news for people running businesses, every weekday, with a VarOps columnist joining us on one story. Today, a real-world test of AI coding agents, Microsoft getting choosy about where its AI runs, and new lawsuits over AI voice and music clones. Rex Factor, who writes Proof of Work, joins us on the coding agents. Plus four quick hits, and what each one means for you.Today's stories00:45 Alibaba's Co-Design Bench, built from real user coding sessions, finds the best agent (GPT-6 Astra) meets about 70% of each task's checklist but fully completes only 26.4% of tasks, and under 18% of codebase-wide changes. Rex Factor, who writes Proof of Work, on treating agent output as a first draft, writing down what done means, and having one person sign off before it touches real data. (arXiv, The Decoder)02:47 Bloomberg reports Microsoft declined to renew its 2023 deal to rent AI computing from Oracle, which ran on older Nvidia chips, and CNBC says it plans to trim its smaller cloud partners. If a plan depends on an AI feature arriving by a set date, plan around what's generally available today. (Reuters via Investing.com, CNBC)04:10 Rights holders for Elvis Presley and an Eminem publisher are suing Suno, and the Presley suit cites Tennessee's ELVIS Act on voice cloning. If your company uses AI voices or music, check the tool's commercial terms and never prompt for a real artist's style or voice in public work. (The Verge)Quick hitsApple plans finer Siri control of its own apps in iOS 27.3, and actions inside third-party apps likely in the spring (Bloomberg's Mark Gurman). (The Verge)Anthropic reportedly takes over about $3.4 billion of gas generator orders, 1.7 gigawatts, for its data centers (Financial Times). (CNBC)Google Meet admins can show custom policy text before calls, for example on recordings, transcripts and Gemini notes. (Google Workspace Updates)Nvidia's board opens an independent investigation after reports CEO Jensen Huang was targeted by a foreign intelligence effort; it says it has no reason to believe secrets were compromised. (NVIDIA)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

The week in AI for business leaders. Ramp's new data says AI spending has flattened even as usage climbs, so we look back at Revolut, Menlo Ventures and Apple's caps with that in mind. Then agents: Uber's four agent rides, Salesforce's Coworker claim and Perplexity's Teammates, next to a real project that counted every result. And what your AI tools can reach, from Postman and Claude Code to exposed AI hacking tools and Conductor's shutdown. North Wayne, who writes First Opinion, and Rex Factor, who writes Proof of Work, join us. Plus what to watch next week.This week's stories00:29 Ramp's data shows AI token spend grew just 1.5% across August and September while usage kept rising, and the median company spends $12 to $15 per employee a month. We revisit Revolut running its own model, Menlo's budget overruns and Apple's daily caps, plus OpenAI's $11 billion raise. North Wayne, who writes First Opinion, on using a buyer's market at renewal time and keeping caps people can ask to lift. (TechCrunch (Ramp), The Register (Revolut), Menlo Ventures, TechCrunch (Apple), TechCrunch (OpenAI))04:24 Uber saw just four rides booked by AI agents in its busiest week, Salesforce says Agentforce Coworker passed 185,000 users, and Perplexity's Teammates cost $200 a month. Rex Factor, who writes Proof of Work, on DoorDash's feature flag agents (45 of 50 done, about $4.79 each) and why counting finished jobs beats a user count. (VarOps, Proof of Work, TechCrunch (Uber), Salesforce, The Verge (Perplexity), Mistral)07:24 Claude Code's email fix still hadn't shipped by Saturday morning, the Palo Alto and Clutch deal is still only talks, and Censys found more than 800 Artex AI hacking tool deployments open on the internet. Make a list of every AI tool with access, update Claude Code, turn on multi factor login, and export from Conductor Cloud before it goes read only on 13 November and starts deleting data on 20 November. (GitHub issue, TechCrunch (Artex), TechCrunch (Clutch), Conductor)Watch next weekAnthropic's fix removing the account email from Claude Code's context, promised for early next week. (Hacker News)Anthropic's IPO: possibly as soon as the week before Thanksgiving (WSJ), with some buyers eyeing a much lower valuation (Bloomberg). (TechCrunch)Mistral's new managed agents and apps on your own servers or private cloud, planned for late November. (Mistral)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Perplexity launches always-on AI Teammates at $200 a month, Claude Code was caught putting users' email addresses into web requests, and Palo Alto Networks is reportedly in talks to buy agent security startup Clutch for about $500 million. Gritt Scott, who writes Skill Issue, joins us on Perplexity. Plus four quick hits, and what each one means for you.Today's stories00:21 Perplexity's new always-on Teammates take separate jobs, run around the clock and work in Slack, but need the $200 a month Max plan. Gritt Scott, who writes Skill Issue, on testing one recurring job before paying for seats, and checking what it can read in Slack. (The Verge)02:29 A GitHub report says Claude Code put signed-in users' email addresses into User-Agent headers sent to third-party sites. Anthropic says a fix lands early next week; update to 2.1.317 or later now, and treat anything in an agent's context as something that can leave. (GitHub issue)03:49 TechCrunch reports Palo Alto Networks is in talks to buy Clutch Security, which secures AI agents' access to company systems, for about $500 million, roughly 50 times revenue. Start with a list of every agent and AI tool that can log into your systems. (TechCrunch)Quick hitsClaude Personal Finance connects bank accounts through Plaid, read only, in US beta. (TechCrunch)OpenAI releases GPT-6 Cyber to select Daybreak Red members for defensive security work. (The Verge, OpenAI)Airbnb says its own AI travel agent launches in the first half of 2027. (TechCrunch)Box builds Nvidia Nemotron embeddings into Box Content Retrieval for agents, early next year. (TechCrunch)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Uber says only four rides were booked through an AI agent in Halloween week, Apple is cutting daily internal AI spending caps, and Salesforce claims Agentforce Coworker has 185,000 users at $50 a seat. Nix Nullty, who writes Overhyped, joins us on Uber. Plus four quick hits, and what each one means for you.Today's stories00:21 Uber said only four rides and about 500 Uber Eats orders came through AI agents in Halloween week, out of about 320 million trips. Nix Nullty, who writes Overhyped, on why agent commerce demos still outrun the receipts. (TechCrunch, The Verge)03:01 Bloomberg reports Apple cut some engineering daily AI caps from $300 to $200, with further cuts toward $100, and non-engineering possibly as low as $25. Apple declined comment. Set your own hard limit and a path for exceptions. (TechCrunch)04:33 Salesforce says Agentforce Coworker passed 185,000 users with 21x usage growth in three months, at $50 per user per month. Vendor numbers; pilot one measurable job before buying company-wide seats. (Salesforce)Quick hitsMeta acquires Conductor; cloud service shuts November 20 then deletes cloud data (TechCrunch)OpenAI to acquire Pierre Computer Company for Codex code review tooling (OpenAI)Google: measure cost per successful task; route easy work with cheap classifiers (Google Developers Blog)OpenAI Images 2.6 brings faster generation and GIFs to ChatGPT and the API (OpenAI)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations

Revolut cuts its AI coding bill by about 90% by self-hosting an open-weight model, Menlo Ventures puts enterprise AI spending at $177B, and Postman installs plugins into developers' AI coding agents without asking. North Wayne, who writes First Opinion, joins us on Revolut. Plus four quick hits, and what each one means for you.Today's stories00:20 Revolut cuts its AI coding bill by about 90% by self-hosting an open-weight model for about 2,000 engineers. North Wayne, who writes First Opinion, on what the number leaves out and why it's still worth asking your vendors about. (The Register, AI Engineer)03:20 Menlo Ventures puts enterprise AI spending at $177B this year, and says companies spent 2.5 times what they budgeted. Menlo invests in Anthropic, so read it as a direction, then put your own spend next to your budget. (Menlo Ventures)05:00 Postman installed plugins into developers' Codex and Claude Code without asking, and says it will make the opt-in clearer. Find out what plugins and standing instructions your AI coding tools have, and who added them. (GitHub, RuntimeWire)Quick hitsPalantir's US commercial revenue passes US government revenue, on shorter contracts (TechCrunch)Salesforce agrees to buy Cresta, maker of voice agents for customer service (Salesforce)Google Meet admins can show a policy notice before Gemini records or takes notes (Google)Anthropic publishes FlashProve research on proving which model answered (Anthropic)---AI. For the people who can’t afford to guessAI Squawk is a daily AI news briefing for business leaders, produced by VarOps, the resident AI company. For more in-depth reads built on real data and hard-earned lessons from the trenches, sign up for the VarOps newsletters at VarOps.com ›More from VarOps:📚 Books🎙️ Podcasts🤖 Resident AI🔬 Research🤝 AI Transformations