On Thursday, November 5, the stocks behind the AI boom pulled the whole market down. The S&P 500, the index most retirement funds follow, fell 4.4%, its biggest one-day drop since April 2025. The tech-heavy Nasdaq fell 5.5%. Reuters reports that the 100 biggest companies on the Nasdaq lost about $2.5 trillion in value in a single day. Spread across everyone in the US, that’s roughly $7,400 each.
The companies that make AI’s chips and run its data centers fell hardest. AMD dropped 17.3%, CoreWeave, which rents out AI computing power, fell 9.8%, and Nvidia, whose chips power most AI, fell 9.3%.
One-day change in share price or index level, Nov 5 2026. Source: AP. Bars share one axis.
We covered a smaller drop in Issue #14, when investors flinched at Big Tech’s building bill. This one was more than twice as deep, and the people who watch markets for a living pointed to a different worry: that building may slow. UBS, the Swiss bank, estimates about 30% of the new US data-center capacity expected in 2028 could be at risk from construction pauses. Qualcomm, which makes phone chips, also gave a disappointing forecast. "The market’s moving from the promise of AI to the price of AI," said Ted Parkhill of Incline Investment Management.
Then most of it came back. After a strong Monday, the S&P 500 closed November 9 less than half a percent below where it stood on October 30. Ed Yardeni, a longtime Wall Street strategist, said he didn’t think the AI bubble was bursting, calling it "an AI bubble losing some air very, very rapidly."
Six giants: Nvidia, Microsoft, Amazon, Alphabet, Broadcom, Meta; our choice of which count as AI-heavy. Holdings: State Street SPY, Nov 9. Price change only, dividends excluded.
If you have a 401(k) or an index fund, about $26 of every $100 in an S&P 500 fund now sits in six companies betting heavily on AI. That’s why a bad day for AI is now a bad day for ordinary savers. A week like this is a reason to know what you own, not a signal to sell. Nobody, including us, knows whether this was a wobble or the start of something bigger.
AP closes: S&P 500 Tue Nov 3 7,568.92 (+0.4%), Wed 7,688.62 (+1.6%), Thu 7,353.73 (-4.4%, -334.89), Fri 7,376.86 (+0.3%); week -2.7%. Nasdaq Thu -5.5% to 24,315. CNBC: Mon Nov 9 S&P 7,549.94 (+2.35%), Nasdaq 25,253.14 (+2.74%). Oct 30 S&P close 7,580.07 derived from AP’s Nov 2 figures. Reuters: Nasdaq 100 -6.4%, about $2.5 trillion of market value; NBC puts it at -6.3% and $2.3 trillion. Per-person yardstick uses roughly 340 million US residents. CNBC: Nvidia lost almost $450 billion in value; Qualcomm fell more than 9% after its forecast. Reuters: Philadelphia semiconductor index -12%. UBS estimate via Reuters. SPY weights Nov 9: Nvidia 6.64%, Microsoft 6.21%, Amazon 4.02%, Alphabet 4.69% (two share classes), Broadcom 2.33%, Meta 2.17%, sum 26.06%. Some coverage also cited Election Day results as a factor; we leave that out.
Anthropic released a faster, cheaper model Frontier race
On November 9 Anthropic, the maker of Claude, released Claude Fable 5.2, the newest version of its top public model. The company says it answers up to 30% faster than the last version, and businesses pay $8 for every million tokens they feed in (a token is about three-quarters of a word). Fortune reports that’s cheaper than OpenAI’s top model, GPT-6 Astra, at $10. People on paid Claude plans such as Pro and Team can use it now, within usage limits.
Anthropic’s own tests put Fable 5.2 ahead of OpenAI’s and Google’s best on most coding tasks, though Google’s Gemini 4 Carbon still beat it on one test of routine office automation. The headline claim is about maths: Anthropic says 1,000 copies of the AI working together turned the proof of Fermat’s Last Theorem, a famous problem solved by a human in the 1990s, into a form a computer can check, in 30 hours. An earlier, unreleased research version of Claude took 11 days.
Meta also released Muse Ember, a small, very cheap model built for simple jobs done at huge volume.
The price war matters more than the scoreboard. When the top AI gets cheaper every few weeks, more of the apps you already use will quietly start running on it. Treat the maths claim with care: checking a proof that already exists is impressive engineering, but it isn’t new mathematics, and outside testers did not review Anthropic’s announcement.
Anthropic lists Fable 5.2 at $8 per million input tokens and $40 per million output; Fortune puts GPT-6 Astra at $10/$50. Anthropic-run FrontierCode 2.0: Fable 5.2 49.1%, Opus 5.5 45.5%, GPT-6 Astra 42.8%, Gemini 4 Carbon 41.1%; AutomationBench: Carbon 53.9% vs Fable 51.4%. Mythos 5.2 is limited to vetted cybersecurity and life-sciences organisations. Pro and Team plans get Fable for up to 50% of weekly usage limits. Anthropic’s footnote says UK AISI and METR did not review the post or system card. Anthropic also links two black-hole physics preprints it says the models solved; they are not peer-reviewed and we don’t report them as solved. Muse Ember: $0.10 per million input tokens (TechCrunch). Disclosure: Human Terms is written with help from Claude.
OpenAI’s security model found 948 Linux flaws Nov 5
Linux runs most of the world’s web servers. OpenAI says a special security version of its AI, GPT-6 Cyber, searched Linux for flaws that would let an attacker take full control of a machine, and found 948 of them in one run that would have cost about $1.2 million at OpenAI’s prices, roughly $1,300 per flaw. Ten days in, 61 had been fixed and 43 more had fixes waiting.
In July, an earlier model found 166. OpenAI is giving GPT-6 Cyber only to approved security teams.
Flaws that could give an attacker full control. Source: OpenAI, via Techzine. Bars share one axis.
This is the double edge of AI in security. The same tool that finds holes for defenders could find them for attackers, which is why access is restricted. For you, the practical step hasn’t changed: install updates when your devices ask, because fixes like these arrive that way.
Counts via Techzine (OpenAI’s research post returned an error to our checker): 948 unique root exploits from 2,308 candidates; 61 fixed upstream and 43 with a staging patch ten days after the scan began; July scan with GPT-5.6 Sol found 166. OpenAI’s deployment page: GPT-6 Cyber is "available only to approved Daybreak Red users and organizations" and rated "Critical in Cybersecurity."
Google will turn off AI for K-12 students by default Nov 3
On November 3 Google said its AI tools will be "entirely off" for student accounts from kindergarten through 12th grade until school administrators choose to switch them on. Google says it will give schools details "in the coming months," before the change rolls out. Google Search is on the list too, so it will also be off for students until the school switches it on.
If your child has a school Google account, whether they can use Gemini will now be the school’s decision, not Google’s default. It’s a good question for the next parent night: has the school turned it on, and what are the rules?
Google’s post (Chris Phillips): "Access to these products will be entirely off for K-12 student accounts until school administrators explicitly turn them on." No firm start date given.
Apple starts selling extra AI usage Nov 5
Apple’s Creator Studio apps, starting with Logic Pro, include a monthly allowance of AI features. Hit the limit and you can now buy a one-time $49.99 "intelligence usage pack," 9to5Mac reports. Apple’s support page says unused usage rolls over for up to 12 months and the packs can’t be shared with family.
This is a sign of where AI pricing is heading: a monthly allowance, then a top-up, like phone data. Apple hasn’t said how much use a pack buys. If metered AI spreads to the apps you use, check the allowance before you subscribe.
Apple support article 127888 (Nov 5): packs let subscribers "continue using these intelligence features after reaching the monthly usage limit," "don’t support Family Sharing," and unused usage expires 12 months after purchase. Available in eight countries. Price per 9to5Mac (Nov 9).
Chatbots pointed patients to stem-cell clinics
Researchers at Simon Fraser University in Canada asked six popular chatbots, including ChatGPT, Gemini, Claude and Grok, 540 questions about stem-cell treatments for Alzheimer’s, autism and cerebral palsy, posing as patients in five cities. The chatbots cited about 5,000 links. Nearly a quarter of them, 23.4%, led to businesses selling treatments, not to health agencies or hospitals.
Source: Simon Fraser University, Nov 5 2026 (Connections Digital Health & AI, BMJ).
Some went further than the question. Of the links to businesses, 17.3% pointed to clinics offering a behavioural therapy for autism, not stem cells, which the researchers never asked about. "There’s a big difference between Health Canada and a business that is trying to sell you a product," said Jeremy Snyder, the study’s lead author.
When you ask a chatbot a health question, look at where its links go before you trust the answer. A clinic’s website is an advert. Stem-cell treatments for these conditions are largely unproven, and they’re exactly where sellers look for worried families.
Published in Connections Digital Health & AI (BMJ); release Nov 5. Six chatbots (ChatGPT, Claude, Llama, Gemini, DeepSeek, Grok); cities: London, Montreal, New York, Paris, Toronto; 5,025 links. The journal page was not reachable to our checker, so figures are from the university release.
AI drafts cut heart-scan reporting time
Randomized trial, 1,200 echocardiograms, Kaiser Permanente; AHA Scientific Sessions, Nov 9 2026. Not yet peer-reviewed.
At the American Heart Association’s meeting on November 9, Kaiser Permanente researchers reported a trial of 1,200 heart ultrasounds. With an AI draft, cardiologists took 4.6 minutes to finish a report instead of 8.5, and AI drafts needed big changes about as often as human ones. The results are early and not yet peer-reviewed.
Heart specialists are scarce, so faster reports could mean shorter waits for results. A cardiologist still signs off every one. "AI cannot, however, replace the human team," said Keane K. Lee, one of the study’s senior authors.
1,200 echocardiograms from 21 centers; reviewers blinded to the drafter. Substantial edits needed: 50.6% of AI drafts vs 49.7% of sonographer or cardiologist drafts. About 46% time saved. Late-breaking science, not peer-reviewed; the AI product is not named.
Is AI speeding up its own progress?
For years, some in the AI world have argued that one day AI would start improving itself, and progress would run away. On November 5 METR, an independent group that tests AI, published the most careful outside measurement yet. After eleven weeks working with Anthropic’s data, it estimates the company is improving its AI about 1.9 times as fast as it did in 2024 and 2025, and that AI itself accounts for most of the speed-up.
Source: METR assessment, Nov 5 2026. Estimates, with wide uncertainty. Bars share one axis, 0 to 3x.
METR is careful about how sure it is. Its range runs from 1.5 to 2.8 times, and it says it "would not be surprised if the true acceleration was closer to 1.3x or 3x." Anthropic redacted most of the evidence METR found most useful, so readers have to take METR’s judgment on trust. Measured across all the leading AI companies, the speed-up is smaller: about 1.4 times.
Skeptics say a company with an IPO coming has every reason to look fast. Supporters say this is an outside group with access no one else has had. Either way, it’s worth watching. If the pace you plan your job around is itself speeding up, the gap between announcements will keep shrinking, and that’s the bigger story behind this week’s new models.
METR assessment, fieldwork Aug 12 to Oct 29: 1.9x the 2024-25 rate (95% CI 1.5-2.8x); roughly 80% of the speed-up attributed to AI, flagged as highly uncertain; 1.4x across frontier labs. Anthropic redacted "a majority of the evidence we found most informative"; METR had no query access to unreleased models; Claude helped analyse the data. We covered METR’s arrival inside the labs in Issue #13. Disclosure: Human Terms is written with help from Claude.
Anthropic sets a price range for its IPO
Last week we covered Anthropic’s filing to sell shares on the Nasdaq. This week the price became the problem. On November 6 Reuters reported that potential investors doubted the company could start its planned roadshow, the tour where a company pitches its shares to big investors, in the wake of the sell-off. Some saw a value of $1.5 trillion as the high end, after earlier talk of up to $2 trillion.
Then, early on November 10, Anthropic filed a price range with the SEC: $70 to $78 a share for 825 million shares, about $64 billion at the top of the range, plus $26 billion more from a private sale alongside it. By our math, counting the shares the filing says will exist afterward, that values Anthropic at roughly $1 trillion at the top of the range.
An IPO is when the public, not just private investors, gets to put a price on a company. The range Anthropic filed puts it near $1 trillion, well below the $2 trillion once talked about. Where the shares actually price, and trade on day one, will be a direct read on how much faith big money still has in AI after last week.
Reuters (Nov 6): investors "raising doubts about whether the artificial intelligence startup can proceed with its planned roadshow next week"; the offering could be postponed "until the market looks more resilient"; Anthropic declined to comment. SEC EDGAR: S-1/A (Amendment No. 2) accepted Nov 10, 6:29am ET; estimated range $70.00-$78.00; 801,351,354 shares from Anthropic and 23,648,646 from selling stockholders; $26.0 billion concurrent private placement of Class N stock at the IPO price; Nasdaq symbol ANT. Valuation is our arithmetic: about 13.09 billion Class A and Class N shares outstanding after the offering, times $78, is about $1.02 trillion (about $0.92 trillion at $70), before options and other dilution. Disclosure: Human Terms is written with help from Claude.
Business spending on AI grew about fivefold
Menlo Ventures, a venture capital firm, surveyed about 400 US business decision-makers and estimates that company spending on AI went from $33 billion last year to $177 billion this year. On average, companies spent 4.7 times what they spent last year, after budgeting for 1.9 times.
Most of that money went to two companies. One caveat up front: Menlo is an investor in Anthropic.
Source: Menlo Ventures enterprise AI report, Nov 3 2026 (survey by Morning Consult, Sept 8-26). Menlo is an Anthropic investor. Bars share one axis, 0 to 100%.
The report says 55% of this money is new spending, not cash moved from existing software or salaries. So far, companies are adding AI on top rather than paying for it by cutting staff. It also explains why investors keep asking when all this spending turns into profit.
Menlo Ventures, State of Generative AI in the Enterprise (Nov 3): deployment spend $33bn (2025) to $176.8bn (2026); reported spending averaged 4.7x last year’s level against a planned 1.9x (median 2.2x vs 1.8x). Anthropic and OpenAI together hold 86% of model spend. Agents account for 80% of tokens on OpenRouter, a Menlo portfolio company. Coding $34.0bn, of which Claude Code and Codex $24.0bn. Disclosure: Human Terms is written with help from Claude.
- Monday, Nov 16The Fed’s hearing on AI, productivity and jobs, with outside voices invited.
- Tuesday, Nov 17Nvidia, the company whose chips power most AI, reports earnings after the market closes. It’s the first big test since last week’s sell-off.
- Around Nov 19-20Anthropic’s expected first day of trading on the Nasdaq, at $70 to $78 a share if the range it filed holds.
Transparency note: Human Terms is written with help from Claude, made by Anthropic. Anthropic appears four times in this issue: its new model, an outside audit of its progress, its IPO, and its share of business AI spending. The model claims are Anthropic’s own, the audit is by an independent group, and the spending report comes from one of its investors, which we say where it runs.