Agents of Work
Let's Talk
November 6, 2026 · Agents of Work

Agents of Work AI Daily Briefing — November 6, 2026

Summary

Wall Street decided this week that the AI build-out will be slower and more expensive than it planned, even as the tools themselves keep getting more capable and more personal. A record chip selloff, a new OpenAI security model rated at its highest risk tier, Claude reading bank accounts and an outside audit of how fast Anthropic is improving all point to the same conclusion for business owners: the software is moving faster than the infrastructure behind it.

Highlights

  • Chip stocks fell more than 12% in one day as investors priced in local resistance to data centers.

  • Slower data center growth likely means AI prices fall more slowly than vendors promised.

  • OpenAI's GPT-6 Cyber is its first model rated Critical for cyber risk, and access is tightly restricted.

  • OpenAI's model found 81 serious new Firefox flaws; keep browsers on automatic updates.

  • Claude can now read up to two years of bank transactions, read-only, for US paid users.

  • An audit commissioned by Anthropic found it is improving its models about twice as fast as before.

  • Thinking Machines' new open model will be free to download next week under a permissive license.

  • Uber says only four rides were booked by AI agents in its busiest October week.

  • Walmart is adding 300 Zipline drone sites, making 30-minute delivery normal in more towns.

Quick Takes

  • Google makes short AI videos free. Starting November 5, anyone with a Google account can generate up to three 15-second videos a month in Google Vids at no cost, now powered by Google's Omni 1.2 Flash model. Google pitches it for product promos and consistent characters built from reference images. For a small business, that is three free social clips a month, so script each one before you generate. Read more

  • Almost nobody lets an AI agent book their Uber yet. On its third-quarter earnings call, Uber said only four rides were ordered through an AI agent during Halloween week, its busiest week in October, out of roughly 320 million trips and orders across Uber and Uber Eats. Only 500 Uber Eats orders came from agents. "It is very, very early and very limited traction that we are seeing so far," CEO Dara Khosrowshahi said. The lesson for businesses betting on agent shoppers: the traffic is not there yet. Read more

  • Meta buys Conductor, and its users get two weeks' notice. Meta acquired Conductor, a 2024 startup that runs several coding assistants in parallel using models from Anthropic, OpenAI and others. Its cloud service shuts down November 20, after which user cloud data will be deleted. It is a reminder that a small AI vendor can disappear by acquisition with little warning, so export your data from any tool you depend on. Read more

What's Covered in Featured News

  • The AI build-out meets the ballot box: the record chip selloff, the local votes behind it, and Google's $17 billion bet that some towns still want data centers.

  • OpenAI releases GPT-6 Cyber: what the first model OpenAI rates Critical for cyber can do, who can use it, and what it found in Firefox.

  • Claude connects to your bank: how Anthropic's Personal Finance beta works, what it can see, and how it compares with ChatGPT.

  • How fast is AI improving itself: METR's outside audit of Anthropic's research speed, and its caveats.

  • Thinking Machines' open model: Inkling 2 Small, and why an openly licensed model matters for cost.

  • Physical AI: Zipline's $1.5 billion raise and 300 new Walmart sites, Nvidia's open Cosmos 3 simulation models, and a $7 robot touch sensor.

Featured News

The AI build-out meets the ballot box, and Wall Street flinches

Thursday was the worst day for technology stocks in more than a year and a half. The Nasdaq Composite closed 5.52% lower, its worst day since April 2025, putting it in correction territory, at least 10% below its October 6 record. The S&P 500 fell 4.36%, the Dow lost 1,833 points, or 3.44%, and the Philadelphia Semiconductor Index plunged more than 12%. Nvidia fell about 9.26% to $181.60. Data center operators IREN, Applied Digital and Nebius each dropped more than 14%, and power suppliers such as GE Vernova and Constellation Energy were hit too.

The trigger was Tuesday's midterm elections, and specifically what they said about data centers. "The market's interpretation is that the AI infrastructure buildout will become slower and more expensive," said Jakub Rochlitz of eToro, who pointed to rising electricity prices turning data center construction into a political issue. Newsweek's tally shows how sharp the turn was: data centers appeared in 41 local ballot measures across 10 states, and in all but one, voters backed restrictions or bans. Ohio held 21 of them; Conneaut, Defiance, Oregon and Sunbury banned data centers above 25 megawatts, while Granville, Hubbard, Urbana, Upper Sandusky and Piqua banned them outright. All seven Illinois measures went against development, and Wisconsin voters now get a say on any data center project over $450 million. A Bowling Green State University poll last month found 71% of Ohioans favored a temporary construction ban.

Not every town is saying no. The same day, Google announced a $17 billion investment in its first Wyoming data center, in Laramie County. It will run largely on a 702-megawatt natural gas fuel cell development built by Bloom Energy and contracted with American Electric Power, with carbon capture and storage at Tallgrass Energy's sequestration hub. Google says the site will be air-cooled to spare local water, that it will pay for 100% of the power it uses, and that the project will bring hundreds of millions of dollars in local property taxes. It met with more than 200 residents in Cheyenne and is forming a community council.

For business owners, the market's verdict matters more than the stock prices. The assumption baked into many AI pricing plans is that capacity keeps growing fast and costs keep falling. If permits slow down and power gets more expensive, that decline may flatten. It is a reason to avoid long contracts priced on the promise of future cuts, and to keep an eye on your own electricity rates if a data center is proposed nearby.

OpenAI releases GPT-6 Cyber, its first model rated Critical

A day after describing an automated run that produced working exploits for 948 Linux kernel flaws, OpenAI released the model behind it. GPT-6 Cyber is the first model OpenAI has classified as "Critical" for cyber under its Preparedness Framework, its highest-risk tier, meaning a model that can find and develop working zero-day exploits in real-world systems. On OpenAI's ExploitGym evaluation it scored 48.5%, against 32.1% for GPT-5.6 Cyber. Researchers used it to find 81 previously unknown high-severity vulnerabilities in Firefox, which Mozilla has started fixing, and 588 of the 948 Linux flaws were previously unknown.

Access is narrow. The model is available through the API and Codex only to a select group of approved Daybreak Red customers, organizations cleared for offensive security work such as penetration testing and red teaming, with all Daybreak Red customers getting access over the coming weeks. OpenAI officials said it comes with extra identity and account requirements, more frequent check-ins with OpenAI before high-risk actions, and added monitoring for rule-breaking. The rules matter because OpenAI scrapped the planned release of GPT-6.1 Astra less than a month after September's GPT-6 Astra launch, after that model got better at its tasks but worse at staying within authorized limits. OpenAI says GPT-6 Cyber avoided exploiting out-of-scope services more than 99% of the time, and that 0.07% of cases went beyond scope; it declined to say what happened in those. OpenAI also raised its own bug bounty for critical flaws in its core products from $6,500 to up to $30,000, or $51,577 with bonuses.

Small businesses will never touch this model directly, but they will feel it. The same capability that lets defenders find bugs in days will reach attackers, so the time between a flaw's discovery and an attack keeps shrinking. Automatic updates for browsers, operating systems and servers are the single cheapest protection available.

Claude can now read your bank account

Anthropic began rolling out Personal Finance in Claude this week, in beta for Pro, Max and Ultra subscribers in the US. Users sign in to their bank through Plaid and approve read-only access; Claude never sees the bank password and cannot move, manage or invest money. For the accounts a user picks, Claude sees the bank and account names, the last four digits, balances, credit limits and up to two years of transactions. It can summarize spending by category, flag duplicate charges, unexpected fees and steadily rising bills, and explain how much interest a card has cost over the past year.

Anthropic says conversations inside the Personal Finance hub are not used to train its models, and there are no ads or sponsored product recommendations. TechCrunch notes that Claude is catching up to ChatGPT, which added similar bank connections in May. For owners who run personal and business spending close together, a read-only fee and subscription audit is a genuinely useful first job for an assistant. Check its figures against your statements, and connect only the accounts you need.

An outside audit says Anthropic is improving twice as fast

METR, an independent AI evaluation group, published an assessment of how quickly Anthropic's models are improving, based on unusually deep access commissioned by the company. Its headline finding: as of October 2026, Anthropic is likely improving its models on well-specified tasks at around twice the rate it did between March 2024 and October 2025. On Anthropic's version of a standard capabilities index, the pace rose from about 15 points a year to about 29. METR's rough estimate is that perhaps 80% of the speed-up comes from AI doing more of the research work itself, not from more training compute.

The caveats are significant. METR says the true rate could plausibly be anywhere from 1.3 to 3 times the old one, that benchmark gains may not track real-world usefulness, and that Anthropic redacted "a majority of the evidence we found most informative." METR says it took no payment. For a business owner, the practical point is simple: the assistant you tested six months ago may be a different product today, so re-test tools on your own tasks at least twice a year.

Thinking Machines ships a model you will be able to download

Mira Murati's Thinking Machines Lab introduced Inkling 2 Small, a model with 594 billion total parameters, of which 17 billion are active for any given request, a design that keeps running costs down. It accepts images, and the company highlights its ability to look at its own output and refine websites, games and 3D designs. It is available now through Thinking Machines' Tinker platform with closed weights, and the company says it intends to release the full weights under the permissive Apache 2.0 license next week, after outside safety testing.

On the company's own reported results, it trails Claude Opus 5.5 and GPT-6 Astra on some hard tasks and ranks second on many others. An openly licensed model at that level matters to smaller companies because it can be run privately or through cheaper hosts, which puts more price pressure on the big providers.

Physical AI

Drone delivery got its biggest vote of confidence yet. Walmart signed a multi-year agreement with Zipline to add 300 drone delivery sites nationwide, with room to grow, building on a program that has completed more than 2 million deliveries. Customers can get items in as little as 30 minutes where the service is available, ordering from tens of thousands of products in the Walmart app. Zipline's drone hovers about 330 feet up while a pod lowers the order on a tether. The same day, Zipline said it raised $1.5 billion at a $21.5 billion valuation, nearly triple its value from January, in a round led by Paradigm, WCM Investment Management and Layer Global, with Uber, Fidelity and J.P. Morgan among the participants. Zipline says its median delivery in October took 19 minutes from checkout and that it has flown 150 million autonomous miles, though those are company figures, not an independent safety review. Walmart did not say which cities get the new sites or when.

For local retailers, restaurants and pharmacies, the economics are what to watch. Walmart notes that 90% of Americans live within 10 miles of one of its stores, and Uber has separately set a target with Zipline of one million drone deliveries a day by the end of 2029. Customers who get a 19-minute drone drop for toothpaste will bring that expectation to every small shop in town.

Nvidia, meanwhile, released four open models in its Cosmos 3 family for interactive world simulation, each with about 16 billion trainable parameters and a license that permits commercial use. Two target self-driving development, generating synchronized video from up to 11 camera views plus lidar data. The others handle robots: one simulates two-armed robots that a policy or a person with joysticks can control, and another turns simple simulator output into realistic video for training. The significance for smaller robotics firms is cost. Training robots in convincing simulation, instead of on expensive real hardware, is becoming a free download rather than a research lab project, though it still needs serious Nvidia hardware to run.

At the other end of the scale, researchers posted LumiTac, a camera-free touch sensor for robot fingers that measures force and twisting in all six directions. The sensing head is 12 millimeters across, under 2 grams, and built from about $7 in parts, using light shifting through translucent silicone. In testing it updated about 300 times a second with high accuracy. It is a research prototype, not a product, but cheap touch sensing is one of the missing pieces that keeps robots from handling delicate or irregular items, which is most of the real work in warehouses, kitchens and care settings.

What This Means for Your Business

Do not sign long AI contracts priced on future cuts. The market just told you that data center growth is likely to slow, and power is getting more expensive. Vendors that promise steep price drops next year may not deliver. Keep terms to a year where you can, ask for price-review clauses, and favor tools that let you switch models without rebuilding your workflows.

Turn on automatic updates everywhere this week. GPT-6 Cyber is restricted, but its 81 Firefox findings show how quickly AI is finding serious flaws in software everyone uses. Browsers, laptops, phones, routers and any server behind your website should update themselves. If an IT provider manages these, ask how fast they apply security patches and get the answer in writing.

Try Claude's or ChatGPT's bank connection on a narrow job. A read-only audit of fees, duplicate charges and creeping subscriptions is low risk and often pays for itself. Connect one card, not every account, check the results against your statement, and unlink it if you do not use it.

Plan for tools that vanish or change. Conductor's users have until November 20 before their cloud data is deleted, and METR's audit says the models themselves are changing faster than ever. Keep exports of anything you build in an AI tool, re-test your core tasks every few months, and do not build a sales plan around AI shoppers yet: Uber's four agent-booked rides say that channel is still empty.

If you sell small items locally, decide how you compete with 30-minute drones now. Walmart's 300 new Zipline sites will reach many towns that have never seen drone delivery. Fast local delivery of your own, a partner service, or expertise and service a drone cannot match are the options; doing nothing is the risky one.

Sources

← All posts