Agents of Work
Let's Talk
October 14, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 14, 2026

Summary

Google put its newest frontier model on sale at half price for six weeks, Anthropic says it has cut theft of its model's reasoning by up to 99%, and American Express wants a signed record of what your shopping agent is allowed to buy. The AI business is shifting from showing off models to building the controls that make them safe to hand money, data and tasks. For a business owner, those controls now matter as much as the model.

Highlights

  • Gemini 4 Argon costs half its list price until November 30; test it on long tasks before the price doubles.

  • Independent tests put Argon first on professional work but behind Claude Opus 5.5 on overall intelligence.

  • An OpenAI engineer allegedly bugged an executive conference room; check who can enter yours.

  • Anthropic says it cut theft of Claude's reasoning by up to 99%, and now charges for blocked requests.

  • Amex wants budget, store and product recorded before any shopping agent gets a payment credential.

  • Character.AI is in bankruptcy; export anything your business keeps inside a consumer AI app.

  • Two-thirds of the Forbes AI 50 bill by usage, so put spending alerts on every AI subscription.

  • Walmart's warehouse robots break down often and cost more to run than people, the Wall Street Journal reports.

  • Robot makers now sell fine-tuning by the demonstration; ask for intervention rates before buying.

Quick Takes

  • Google Play will run apps you describe to an AI, privately. Adults in the United States can now build an app in Google AI Studio and install it on their own Android phone through Google Play with no public listing. Sharing with a small circle is promised in the coming months. Keep customer data out of anything built this way until Google says how private apps are checked. Read more

  • X bundles Grok and Cursor into one subscription. The company now called SpaceX Super Intelligence sells five tiers, from free through Plus at $8 a month, Premium at $30 and Super at $100 to Ultra at $200, each combining Grok chat and its agent, the Cursor coding tool and X Premium features. AI usage caps rise by tier, and the newest Grok models require Premium or higher. Read more

  • Reuters says AI labs sought a say over what outside reviewers publish. Until at least October 5, Anthropic and OpenAI pushed for substantial control over what independent safety evaluators embedded in their companies could make public, according to a Reuters report citing two people familiar with the talks, despite both labs promising last month that reviewers could publish their main findings without company edits. Anthropic says long-standing partners decide their own publication while new firms face a sign-off right during a learning period; OpenAI says it only strips secret details. For buyers, "independently tested" means little until you know who controlled the pen. Read more

  • Stripe on why AI companies charge by the sip. Two-thirds of the Forbes AI 50 use some form of usage-based pricing, Stripe reports from its Monetize conference, because costs vary by job: Replit once charged 25 cents per agent checkpoint while some jobs cost it several dollars to run. Stripe also warns that pricing pages now have two readers, the buyer and the AI helping them evaluate; where pricing was not accessible in one test, AI engines quoted anywhere from $100 to $2,400 per user per month, pulled from forums. Read more

What's Covered in Featured News

  • Gemini 4 Argon goes on sale: what Google's new model costs, who can use it, how it scores on independent tests and where it falls short.

  • The OpenAI bugging case: what prosecutors say an engineer hid behind a conference room television, and how OpenAI found it.

  • Anthropic's anti-theft defenses: how rivals copied Claude's reasoning at industrial scale and what blocking them costs legitimate users.

  • Amex's rules for shopping agents: a framework that records what you authorized before an agent checks out.

  • Character.AI's bankruptcy: how a 20-million-user chatbot company ran out of road, and what happens to the platform next.

  • Physical AI: Walmart's robot troubles, Agility Robotics' path to Nasdaq, Physical Intelligence's robot API, Industrial Next's factory stations and Flox's wildlife deterrents.

Featured News

Gemini 4 Argon goes on sale at half price, with independent tests split on how good it is

Google made Gemini 4 Argon broadly available on October 13, two weeks after unveiling it to a small group of cybersecurity partners. It is now in the paid Gemini API, Google AI Studio, the Antigravity coding tool and eligible Gemini Enterprise plans, and Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search. Koray Kavukcuoglu, who now runs Google DeepMind, wrote that Argon "is just the beginning" of the Gemini 4 family.

The price is the headline for anyone paying for AI by the unit. Through November 30, developers pay $2 per million units of input text and $10 per million of output, with cached input discounted 95%. From December 1 the price doubles to $4 and $20, matching Claude Opus 5.5 and Meta's Muse Wonder. At the introductory rate, Argon costs the same as Claude Sonnet 5.5 and GPT-6.1 Sol. The API accepts up to one million tokens of combined input and output for a single long task, a developer feature rather than how the consumer app behaves.

The independent scorecards disagree. On the Vals Index, which measures finance, coding, legal and tax work, Argon ranked first as of October 12 at 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%. On the broader Artificial Analysis Intelligence Index it scores 53, tied with GPT-6 Astra and Claude Fable 5.1 and behind Opus 5.5 at 58 and Sonnet 5.5 at 56. Sam Witteveen of Red Dragon AI summed it up: "This is a really, really, really good model. But is it the best model out there? I actually don't think it is."

Google's own figures lean on security and real work. On the Gray Swan indirect prompt injection test, where hidden text on a web page tries to hijack an AI agent, attacks succeeded only 0.7% of the time across 15 adaptive attempts. Inside Google, more than 100,000 employees use Argon monthly; over four weeks agent-written code submissions rose 35% while rollbacks fell by nearly a third. Those are company-reported numbers from private tests. The safe reading for a business is that Argon is a top-tier model at a discount for six weeks, strong on professional document work, and not a reason to switch tools you trust without testing on your own jobs.

Prosecutors say an OpenAI engineer bugged an executive conference room

Federal prosecutors have charged Abdulrahman Alfozan, a 35-year-old OpenAI engineer, with using or attempting to use a device to intercept oral communications after a concealed recorder was found in an executive conference room. OpenAI says its technical surveillance team found the device on September 30 during a sweep, plugged into a USB port behind a television and wrapped in toilet paper. Court documents identify it as a small Seeed XIAO ESP32-S3 Sense development board with encrypted storage. On October 5, prosecutors say, Alfozan left a bag in a printer that security staff found to contain his wallet, his ID and a second, similar device.

Alfozan, a Saudi national and U.S. permanent resident, joined OpenAI in July 2025; a LinkedIn profile lists him as an engineering lead on its AI agents platform. He was arrested on October 9, fired the same day in coordination with the FBI, and remains in custody ahead of an October 19 detention hearing. U.S. Attorney Craig Missakian called it "a matter of national security." OpenAI says it used its own models to identify the device and trace its origin, and has no indication model weights, source code or customer data were accessed. The lesson travels beyond AI labs: a hobby board the size of a stamp can record a boardroom, and the people with badge access are the ones who can plant it.

Anthropic says it cut theft of Claude's reasoning by up to 99%

Anthropic published a detailed account on October 12 of how rivals have been copying Claude and what it has done about it. The technique is distillation: an operator sends millions of requests through networks of fraudulent accounts, collects the answers and especially Claude's reasoning transcripts, then trains a cheaper model on them. A February report described one network running more than 20,000 fake accounts at once. Anthropic names Zhipu, which it says distilled cyber capabilities from leading U.S. models while building GLM-5.3, and Moonshot, which it says secretly forwarded requests from its Kimi models to Claude.

The defenses now stack. New API accounts created since August 31 cannot edit prior context in Claude Fable 5.1 and Opus 5.5, a two-stage classifier blocks extraction attempts with a false-positive rate below 0.1%, repeat offenders face stricter limits and bans, and since September 24 Anthropic charges for requests its safeguards block. The company reports a 90% to 99% drop in harvested data, tens of thousands of banned accounts, often within two hours of a campaign starting, and "trillions of tokens" of paying traffic turned away. The cost falls partly on honest customers: some developers had to rework integrations, and a small share of legitimate requests get blocked and billed.

Amex wants proof you approved what your AI agent buys

American Express introduced what it calls an Intent Intelligence Framework on October 13, aimed at the moment an AI shopping agent reaches the checkout. Before any purchase, an API records the customer's instructions and limits, such as the merchant, a maximum budget and product preferences. The customer verifies their identity and approves that record, which produces a Proof of Intent Token used to get a digital payment credential the agent uses instead of the card number. Hannah Bozian, Amex's vice president of agentic partnerships and strategy, gave the example of a customer who asks for white Nike sneakers in size 9 and an agent that tries to buy Adidas: Amex could block it before it completes. Amex says it has not yet seen a rise in agent-related disputes, has no launch date, and is in talks with payments partners. Its own survey found 66% of U.S. consumers likely to use agents to research prices over the next 12 months. If customers' agents start arriving with pre-authorized intent, your product data had better match what they were told to buy.

Character.AI files for bankruptcy five years after launch

Character.AI, the chatbot platform that counted 20 million monthly users in October 2024, filed for Chapter 11 bankruptcy on October 13 with assets and liabilities each between $50 million and $100 million. Disney has agreed to pay $20 million for some assets plus a perpetual, non-exclusive license to the technology. Chief executive Colin Adams said the platform will keep running during a court-supervised sale expected to close in late November or early December, but no bidder has emerged, and if none does, "the platform will be wound down." Founded in late 2021 by ex-Google engineers Noam Shazeer and Daniel De Freitas, the company raised more than $193 million and lost its founders back to Google in a $2.7 billion licensing deal in August 2024. It was sued after a 14-year-old Florida boy died by suicide and eventually banned minors. The filing does not say what happens to user data.

Physical AI

The Wall Street Journal's account of Walmart's warehouse automation, summarized by Supermarket News and Futurism, is the week's most useful robotics reading for any operator weighing a robot purchase. Walmart has spent years and billions automating about 200 U.S. warehouses, and the problems have multiplied: cardboard boxes too big for the machines, wheeled robots that break down often and cannot lift heavy items like bags of dog food, and buildings from the 1980s and 1990s retrofitted around systems from several vendors. The money is not yet working either. Sources told the Journal that the engineers needed to keep automated sites running cost more than the warehouse workers they replace, and electricity at an automated grocery facility can reach $800,000 a month against about $250,000 before. Walmart's response: "New technology is never perfect on day one." Kroger, meanwhile, wrote off $2.6 billion on its own automated network. The deepest pockets in retail are still paying tuition.

Against that backdrop, Agility Robotics cleared the Securities and Exchange Commission's review of its merger with Churchill Capital Corp XI, with shareholders voting November 2 and a Nasdaq listing under the ticker AGLT expected in November. The deal would raise more than $620 million if no shareholders redeem, including a roughly $200 million private placement led by Foxconn. Agility says its Digit robots have logged more than 65,000 hours of real work at customers including Schaeffler, GXO and Toyota Motor Manufacturing Canada, and that Digit 5 enters early access in early 2027 backed by more than $300 million in committed orders, subject to milestones. Its filing, covered here yesterday, showed modest revenue against a large operating loss. Order books, not shipments, are still what the market is pricing.

The tooling for smaller robot makers is maturing fast. Physical Intelligence opened a waitlisted API on October 13 that lets a company upload demonstrations of a task, fine-tune one of its robot models on managed computing and run the result on its own hardware. Early partners include Actor, whose autonomous wheel loader loaded a dump truck at a commercial site in Utah with no operator aboard, and Almond, whose dual-arm Axol robot learned to deburr spacers for its own assembly after four days of iteration. The company reports no success rates or pricing, and notes that cloud-hosted control hit network connectivity problems at some sites, exactly the failure a factory should plan for.

Two Disrupt-stage startups show where that stack lands on the floor. Industrial Next, founded in 2021 by former Tesla engineers Lukas Pankau and Allen Pan, builds self-contained robotic stations taught new tasks through simulation and demonstration rather than weeks of programming; it says one station peels protective film from phone camera modules "at or above human speed" with "100% accuracy," and BMW's technology office is evaluating the system. No funding, pricing or unit counts were disclosed. Flox Intelligence, a 2020 spinout from Sweden's KTH, mounts battery-powered camera devices that identify animals and play species-specific sounds to keep them off runways, rail lines and roads. About 150 devices are live on three continents, and the company expects about $1.5 million in revenue this year. The FAA estimates wildlife strikes cost U.S. aviation about $272 million a year, the kind of specific, measurable cost that makes a robot pay for itself.

What This Means for Your Business

Use the Argon discount as a six-week test window, not a reason to switch. Pick three real tasks your team already does with an AI tool, such as summarizing a contract, drafting a proposal from a project folder or reconciling a spreadsheet, and run them side by side on Argon and whatever you use today. Judge on finished work and errors, not on benchmark charts. If Argon wins, lock in your usage before December 1, when the price doubles; if it does not, you have lost nothing but a few dollars of API spend.

Put spending alerts on every AI tool now. Stripe's data says usage billing is the norm, and the X bundle, Argon's tiered access and Replit's shifting prices all point the same way. Ask each vendor for a hard monthly cap and an alert at 80% of it, and assign one person to read the invoices. If a vendor cannot offer a cap, treat the subscription as a variable cost in your budget, not a fixed one.

Get ahead of agent-driven shopping. Amex's framework means customers' agents will soon arrive with a recorded budget, merchant and product spec. Make sure your product names, sizes, prices and availability are accurate and machine-readable on your site, because a mismatch between what the agent was told to buy and what your listing says will become a blocked sale. Also check whether your pricing page is readable without JavaScript and not blocked to crawlers; if an AI cannot find your price, Stripe's test shows it will guess, badly.

Treat consumer AI apps as temporary. Character.AI had 20 million users and a Disney deal and may still shut down with no word on what happens to user data. If anyone on your team keeps customer notes, scripts or workflows inside a consumer AI product, export them this week and keep the master copy somewhere you control. Prefer business plans with data export and deletion terms in writing.

Audit physical access and meeting rooms the way you audit passwords. The OpenAI case involved an employee with a badge and a board the size of a postage stamp. Limit who can enter rooms where sensitive conversations happen, check behind screens and under tables before a confidential meeting, and make sure departing staff lose badge access the same day. And if a robot vendor is pitching your warehouse or shop floor, ask how often a technician has to intervene, what the electricity bill looks like and who maintains the machines. Walmart's experience says those three numbers decide whether automation saves money.

Sources

← All posts