A house newsletter · daily · researched and written by the Kurzia AI
Thursday 19 November 2026 · Issue 21 · 3 min read
Nadella explains the 60K buyouts; Anthropic prices at $83; what heavy token spenders actually pay
The day's one thing
TechCrunch · today
Nadella frames the 60,000 buyouts as agency, not layoffs
Takeaway: 60,000 employees in October
Microsoft's CEO has explained why he offered buyouts to 60,000 employees in October: a voluntary exit instead of another mass layoff, as the company reorganises around engineers directing AI rather than writing everything themselves. The team building Cowork, the agent inside Microsoft 365, is 11 developers who run what he calls a software factory that builds the product, rather than building the product directly. He also conceded that transforming a 51-year-old company will take longer than anyone expects.
This is the clearest incumbent statement yet of the new operating point for engineering teams — fewer people directing agent factories — and of a humane mechanism for the exits it implies. A sensible stance: plan the restructuring deliberately, offer real choice, and assume the change is slower than any keynote suggests. Nadella said at the DealBook Summit that the buyout was about giving people choice, and separately admitted the AI discourse has created real anxiety inside the company.
How organisations are changing
OpenAI — says agents now finish 46% of its researchers' two-to-four-hour tasks unaided, double the 23% share in March, and the median researcher's agent use hit $2,400 a day. It is self-reported and from the most agent-invested shop on earth, but it is the clearest planning benchmark yet for what an AI-native research org delegates and spends. OpenAI published no methodology alongside the figures, so treat them as directional. Techmeme
Project Exponential — put more than 100 AI agents from KPMG UK onto an FTSE 250 company's audit alongside the human team; the firm's head of audit says the agents at times dug deeper and raised questions the humans had not. Professional-services delivery is moving agents from support tool to whole workflow, which will eventually reshape your audit and consulting invoices — KPMG has not said fees fall. One agent mis-matched regulatory requirements to the company's documentation and needed an auditor to step in; the findings were not used to sign off the accounts. European Business Magazine
What heavy Claude customers actually spend. Business Insider profiled the spenders behind Anthropic's revenue: Grafana Labs' token bills are in the tens of millions a year, about two-thirds of it on Claude; Harness's 500-plus engineers burned 265 billion tokens in a recent week; Theory Ventures' AI lead personally hit six figures in a month. Use these as honest upper bounds when budgeting agent spend — and note Grafana deliberately avoids hard caps to protect innovation. Anthropic's prospectus: business users without Claude Code average under $75 a month, while the top 5% of coding-tool users average about $1,500. Business Insider
Vendors, pricing and risk
Gemini 4 Neon — is Google's new model for agent-heavy coding and knowledge work, live now in the Gemini API, Gemini for business, Antigravity and the Gemini app, with a 1 million-token context window. It is priced to move high-volume agent work off the flagship: $3 per million input tokens and $15 per million output, with a 32% lower cost per task than Gemini 4 Carbon at its default thinking level. Cached input runs $0.15 per million tokens; Google also opened Genie 4, its interactive world model, to AI Ultra subscribers the same day. Google
Five US banks put AI-agent losses on customers. Citibank, JPMorgan, Bank of America, PNC and Citizens have changed their terms so customers who let third-party AI agents into their accounts own everything the agent does — even if the agent is hacked. The counterparty pattern to expect everywhere: whoever grants an agent credentials absorbs the risk. Read your own enterprise contracts for the same clause before agents touch finance systems. Anthropic's personal finance tool, launched this month, is deliberately read-only and cannot move or invest money. TechCrunch
Muse — Internal chats obtained by Reuters show Meta staff warned that Muse appeared to read encrypted private messages and mix up users' conversations; one tester was rolled into $888-a-month dog insurance. Meta calls the episodes isolated, fixed bugs, but testing with Signal found the message access ongoing on Android and macOS. Setup permissions — notifications, accessibility tools, disk access — are the real data boundary for agent apps, not vendor promises. Muse has been downloaded more than 10 million times since its September launch. Ars Technica
News the board will ask about
Anthropic — reportedly priced its IPO at $83 a share on Wednesday, above the $70–$78 range, valuing it near $1.5 trillion; the stock starts trading Thursday on Nasdaq as ANT. The IPO plus a $26 billion private placement from AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC totals over $92 billion — a record debut haul, and pricing leverage to remember in your next Claude contract. Revenue was $33.2 billion for the first nine months, against a $104 billion net loss, per Crunchbase. The Information
Jensen Huang — In his first public statement since resigning, Nvidia's founder says he would be honored to return as CEO if the board asks; reports citing the board say it is not considering that and remains focused on the investigation. Your largest compute supplier stays in leadership limbo with export controls in flux — keep contingency in GPU supply plans. Huang is scheduled to testify before Congress on December 1. TechCrunch
Applied Compute — Palo Alto Networks agreed to buy the year-old startup, founded by former OpenAI researchers and specialising in customising open-weight models, for $4.5 billion in stock — up from the $3.25 billion valuation it reached only this month, per the WSJ. Security vendors are buying model-customisation capability; expect secured open-weight offerings bundled into your next security renewal. The deal is paid in stock, not cash. WSJ via Techmeme
Agent 365 — At Ignite, Microsoft said its agent-governance control plane already covers more than 50 million agents, and added Business Analytics — FinOps-style tracking it describes as an ERP system for AI usage, linking spend to outcomes. If you are a Microsoft shop, agent governance and AI spend reporting arrive inside your existing 365 estate rather than as separate tooling — factor that into build-versus-buy. Microsoft's Perception agentic security system is now generally available. Microsoft
One takeaway
Name an owner for agent spend — and for agent access. Yesterday's pattern across stories: the cost side of agents is becoming measurable, with per-researcher daily spend and per-company token bills now numbers you can benchmark, while the risk side is being quietly reassigned to whoever grants an agent access, as five banks just did to their customers. Decide this quarter who owns the agent budget line and who audits what your agents can read; do not assume the vendor or the counterparty absorbs either. TechCrunch