Anthropic Files for IPO With $16.3B Revenue, $80.7B Loss
ALSO: Microsoft offers 60K buyouts, banks block AI agents
Welcome to Guru99 AI Report!
Top Story: How does a company grow revenue 14x and still post an $80B loss? This week’s report explains, along with Microsoft’s 60,000-person exit offer, an AI diagnostician outscoring doctors, and why your bank may block AI agents.
🚀 Anthropic Files for IPO: $16.3B Revenue, $80.7B Paper Loss
Brief Buzz:
Claude maker Anthropic filed its IPO paperwork with the SEC on October 26. It’s planning a Nasdaq listing under the ticker ANT, with Goldman Sachs and Morgan Stanley as lead bookrunners. OpenAI has ruled out a listing in 2026. Share price and count are still to be determined.
- $16.3B in first-half 2026 revenue — roughly 14x the $1.2B from the same period last year.
- $80.7B net loss, of which $72.4B is a non-cash accounting charge tied to the massive jump in valuation. Operating loss: $693M.
- 11,900+ customers have paid $100K+ over the past year, and existing customers are spending roughly 9x what they did a year ago.
- 48% of revenue comes through the Amazon, Google and Microsoft cloud marketplaces, and the top 5 customers account for around 25%.
- $8.9B spent on compute in the first six months, plus $265.6B in minimum compute commitments running over the next 20 years.
💡 Why Should You Care?
Enterprises are already putting real money behind AI agents like Claude Code, which writes, tests and debugs code. Expect more AI-written code landing on QA desks — and more demand for testers who can test it.
🚨 AI-Assisted Hunt Exposes Critical Flaws in PHP, the Web’s Backbone
Brief Buzz:
Researchers using OpenAI’s GPT-5.6 Sol found two critical bugs in PHP, the language behind around 70% of websites whose server-side technology is known. Patches landed on October 21, and attackers moved quickly to exploit the flaws, so updating is essential.
- Calif found one using GPT-5.6 Sol via OpenAI’s Patch the Planet programme; Wiz found both.
- Rated 9.2/10, the bugs (CVE-2026-109618 and CVE-2026-109619) let an attacker run code on the server without logging in.
- PHP has released patches for versions 8.2 through 8.5. 8.1 and earlier get no official fixes.
- The FBI is urging organizations to patch immediately or take systems offline until the issue is resolved.
- OpenAI says its AI beat attackers to the bugs, but Amazon logged an exploit attempt four days before disclosure.
💡 Why Should You Care?
A clean app-code scan won’t catch this — the bug lives in PHP itself. Check your runtime version, and expect the spike in AI-found bugs that OpenAI is predicting.
💼 Microsoft Opens the Exit Door for 60,000 Employees
Brief Buzz:
Microsoft just offered about 60,000 people a paid way out. Its new Voluntary Exit Offer gives eligible staff financial and career support as they leave. It’s the company’s second buyout scheme this year, and the decision window closes in early December.
- Roughly 27% of Microsoft’s 223,000-strong global workforce is eligible, according to AFP.
- It covers four countries (US, UK, Canada and India) and staff at level 70 and below — roughly up to VP.
- Not every team is in: sales, some marketing and some engineering are sitting this one out, according to KUOW.
- Terms are largely standard severance — tenure pay, stock, health care and career support, a spokesperson said.
- About 30% of eligible staff took the early retirement offer in April. Microsoft expects lower uptake this time.
💡 Why Should You Care?
Microsoft is cutting headcount while pouring money into AI — though it says AI isn’t what’s driving the reductions. For testers, now is the perfect time to build AI-assisted testing skills.
🧮 AI Agents Machine-Check Math’s 15,000-Page “Enormous Theorem”
Brief Buzz:
AI math startup Axiom Math says it has turned the Classification of the Finite Simple Groups — mathematics’ “periodic table of symmetry” — into code a computer can check. Widely regarded as the longest proof ever written, the job took just 24 days.
- The original proof runs some 15,000 pages across 500+ papers by around 100 mathematicians, built up over decades.
- Humans directed dozens of AI agents to write roughly 43.1 million lines of Lean, the language that lets a computer verify every step.
- Most issues were simple typos or table slips, but Axiom logged 34 significant deviations on its public errata list.
- Biggest catch: a false step in a crucial 1986 argument. The theorem still holds — the AI built a new proof from Solomon’s 1976 work.
- Axiom has open-sourced the code and invites experts to re-run the check and confirm the formal statement matches the theorem.
💡 Why Should You Care?
Mathematicians who rely on the theorem can now trust a machine check instead of wading through 15,000 pages. The AI also showed it can spot genuine errors in published math, not just reproduce it.
🩺 Microsoft’s AI Diagnostician Beats Doctors in Peer-Reviewed Study
Brief Buzz:
Microsoft’s MAI-DxO works like a panel of virtual doctors: AI agents propose a diagnosis, order tests and debate before reaching a conclusion. It was tested on SDBench in a peer-reviewed Microsoft paper in Science, where each case unfolds step by step and every test carries a cost.
- 83% vs 36% on 304 hard New England Journal of Medicine cases, against generalists using textbooks and the internet (22% without them).
- 80% vs 47% on 325 past ER admissions from Boston’s Beth Israel Deaconess — records, not live patients.
- ~4 points: MAI-DxO’s lift over plain GPT-5 (78.9%) on the hard cases. On the ER cases, GPT-5.5 scored ~85% with or without it.
- Half to a third of the cost, per Microsoft — but only in cost-capped runs. Its top scorers actually spent more on tests than the doctors did.
- Next up: a live patient study at Beth Israel Deaconess comparing its calls with those of ER and urgent care doctors. No start date yet.
💡 Why Should You Care?
AI second opinions could ease pressure on overstretched ERs. But these were past, text-only cases judged against a single generalist rather than a specialist. There’s no product or pricing yet — Boston is the real test.
🏦 Big Banks Are Locking AI Agents Out of Accounts
Brief Buzz:
AI agents are logging into people’s bank accounts, and banks aren’t happy. An Axios investigation into unvetted AI agents found outside agents inside customer accounts at several major US lenders, some acting beyond what customers asked. Reuters has since confirmed the restrictions.
- Morgan Stanley is blocking outside agents (though it has seen no rogue ones); JPMorgan, BofA and Citi are all restricting them.
- Many agents log in with the customer’s own password — which, even with consent, breaches most banks’ policies.
- Some agents reportedly contacted support lines or went beyond their instructions. No customer losses have been confirmed.
- One bank executive admitted they can’t always tell a user-approved agent from a rogue one.
- Most banks are moving to read-only access — agents can see balances but can’t move money. JPMorgan already runs on that model.
💡 Why Should You Care?
Handing an AI assistant your bank password likely breaks the rules and, as JPMorgan’s Jamie Dimon has said, puts the liability on you. Banks want a middle ground, but JPMorgan has no timeline for it.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
