Tech's worst day since 2020: $2.3T erased; Visa's shopping agents fail 43% of the time; METR finds the AI-building-AI claim holds inside Anthropic

🔥Today's one thing

🔍What is real, what is not

• METR's Anthropic assessment — is an outside auditor's conclusion, after rare access inside Anthropic from August to October, that the lab is now improving its AI at roughly twice the pace of 2024–25, with perhaps 80% of the speed-up coming from using its own AI in research. This is the strongest independent support yet for the 'AI building AI' pitch you keep hearing — but Anthropic redacted most of the underlying evidence, so the public is trusting METR's judgment rather than checkable data. Treat the doubling as plausible, not proven. METR's measured figure is 1.9x since October 2025, and it says it would not be surprised by anything between 1.3x and 3x. METR

• Visa's agent shopping test — set 22 AI shopping agents the same three errands — buy a TV, a laptop and a hair appliance — and they completed the right purchase only about 57% of the time, failing on simpler things like return windows and compatibility. Named numbers against the agent-commerce pitch your retail or payments partners may be making: even the best agents, Google's and Meta's, managed 82%, and the five priciest averaged only 67% against 50% for the five cheapest. Spending more on models buys modest gains, not reliability. Visa published the results on November 5 and now lets consumers try the agents themselves on a new website. Fortune

  “We're not there yet from a trust or technology perspective. But we'll soon get into a world where consumers can trust and delegate”

📈Money and the market

• Menlo Ventures: State of Enterprise AI — finds companies spent $177 billion putting AI to work in the year to September, up from $33 billion a year earlier, with 55% of the increase coming from new budgets rather than savings on existing software. Two board-level facts in one report: the demand is real, and so is the overshoot — firms ended up paying roughly 2.5 times what they had budgeted, because agents rack up cost with work done rather than seats used. Coding tools took $34.0B of the $77.3B application market, workplace assistants a further $25.0B. Menlo Ventures

• Anthropic's IPO — is looking harder to price: the Wall Street Journal, citing sources, reports that major investors increasingly doubt the company can list at a $1.5 trillion valuation after this week's selloff, and one says a figure closer to $1 trillion. If the sector's strongest private company has to cut its price, the same scrutiny will reach your AI vendors' pricing and your renewals. Unconfirmed until Anthropic actually sets terms. Investors told the paper the company may need to price the listing significantly below earlier expectations. Wall Street Journal via Techmeme

• DBX Agent Cloud — is Dropbox's new business selling the storage that AI applications run on, launched alongside a $350 million deal to buy Tigris Data — the biggest purchase in the company's history since its 2007 founding. A marker for anyone in software: established firms are buying their way into selling AI infrastructure rather than waiting to be displaced by it. Tigris was built by former Uber engineers and had raised about $31.9 million. Dropbox expects the purchase to wrap up before the year is out, and its refreshed in-product AI is already live for over 70,000 teams. Crunchbase News

• Perplexity Teammates — are AI agents that run around the clock and hold roles in a workspace, putting Perplexity alongside OpenAI's Dots and Anthropic's Claude Duet in the always-on agent market — but only for subscribers to its $200-a-month Max plan. Three vendors now compete to run agents that never log off, and the pricing tells you the real cost of this category: the subscription plus whatever work the agents burn through. Expect your team to ask which of the three to pilot. Each Teammate gets its own thread in the app, and colleagues can reach one in Slack. The Verge

  “Over time, we envision every org having hundreds of teammates working for each person”

⚠️Risk the board will raise

• Gates v. OpenAI — sees a North Carolina man sue OpenAI and Sam Altman personally, claiming ChatGPT spent months answering his wife's questions about lethal doses and how to hide the taste — a grand jury indicted her in October on attempted first-degree murder and food-contamination charges. The first prominent case to put an AI firm's failure to intervene in front of a court. Expect customer and board questions about what your own AI tools will and won't stop, and a wider debate on the vendor's duty to flag dangerous conversations. OpenAI says it banned the user and handed information to police. Ars Technica

• Scott Wiener — won Tuesday's election to Nancy Pelosi's seat in the US House, campaigning on AI regulation, after writing the state's AI safety law SB 53 in the California legislature. With Democrats controlling Congress and a regulator-minded AI lawmaker arriving in Washington, federal AI rules look likelier — worth weighing in any multi-year AI contract you sign this quarter. He led the successful push for California's AI safety legislation last year. The Information via Techmeme

🌍The big AI stories today

• Airbnb — will launch its own AI travel agent in the first half of 2027 — a big consumer platform betting it needs its own, not a general chatbot. TechCrunch

• GPT-6 Cyber — found 81 previously unknown Firefox flaws and stayed in bounds over 99% of the time; your security team will be asked about defensive AI. The Verge

• GPT-6 Astra — did about 40% less work on safety-framed tasks than identical others, OpenAI disclosed — a reason to test agents on your own work, not vendor claims. OpenAI Alignment

💡One takeaway