AI news, November 9 — Is Opus 5.5 really dumber? Complaints flood in over claim that "something changed on Nov. 3"

Monday, November 9, 2026

7 stories worth reading from 70 on X in the past day

Listen to today's news

Every story, read aloud.

0:000:00

Today's stories

7 stories

3 hours agoBreakingJensen Huang admits affair as Nvidia's board opens probe into foreign spy targetingNvidia CEO Jensen Huang has admitted to an extramarital affair, and the company's board has opened an independent investigation into whether he was the target of a forei…@nytimes +2· ♥ 12K· CNBC
29 hours agoIs Opus 5.5 really dumber? Complaints flood in over claim that "something changed on Nov. 3"Claims that Anthropic's top model, Claude Opus 5.5, has gotten worse since launch have spread on X for days.@AIMind_Ai +3· ♥ 2.8K· X (Shane Melaugh)
18 hours agoCodex Pro plan overhaul angers Plus users; OpenAI's developer account posts one word: "Tomorrow"A member of OpenAI's Codex team said all Pro plans have no 5-hour limit, and Plus subscribers answered with c…@thsottiaux +1· ♥ 8.1K· X (Tibo)
11 hours agoAnthropic secures $3.4 billion of gas generators, "largest known single purchase by an AI lab"Anthropic has agreed to pay $3.4 billion for gas-fired generators to power data centers, the Financial Times…@business· ♥ 564· 지이코노미
30 hours agoMusk: "Grok Bot fixes for token efficiency and smarter behavior underway" as 'nerf' complaints pile upElon Musk said on Nov. 8 that fixes to improve Grok Bot's token efficiency and make it smarter are underway.@elonmusk +1· ♥ 34K· X (Elon Musk)
13 hours ago"Destructively scan all the world's books": Anthropic's 'Project Panama' photos draw 550,000 viewsPhotos from Anthropic's 'Project Panama,' in which it bought millions of books in 2024, stripped the bindings…@yachty66· ♥ 3.4K· X (yachty66)
8 hours agoApple CEO Ternus presses Federighi to "speed up AI road map"; agentic Siri targeted as soon as next yearApple CEO John Ternus held a closed-door meeting in recent weeks with software chief Craig Federighi and push…@markgurman +1· ♥ 2.3K· X (Mark Gurman)

Get AI News Daily in your inbox!

Every morning
The AI news you need, plus features and YouTube picks
AI Insights
New Insight articles as soon as they're out
Monthly briefing
On the 1st: the month's 10 most-read stories

No account needed, unsubscribe anytime in one click.

Trending on YouTube

3 stories

AI videos that drew views and comments fast on YouTube, with the claims checked against articles and sources

AI bubble"AI came for my job, but left with my pension": 'The AI Panic Has Started' video passes 330,000 views
Investor YouTuber Sasha Yanshin's Nov. 8 video 'The AI Panic Has Started' passed 330,000 views within a day.
Per the video description, Blackstone's $41 billion private credit fund limited withdrawals for the second straight quarter after investors asked for $4.3 billion back. The description also claims OpenAI asked for $30 billion and raised $11 billion, Nvidia lost $500 billion in a week, and Anthropic's filing shows it must nearly triple revenue to cover its compute bill. All of these figures are claims in the description.
The top comments were "AI came for my job, but left with my pension" (about 1,000 likes) and "I suggest these investors ask AI how to withdraw their money" (about 1,000). One joke, "Claude, withdraw my money from AI companies, make no mistakes!", got 329. The video carries a disclaimer that it is educational, not financial advice.
- ▶The AI Panic Has StartedSasha Yanshin · 330K views
Jobs"Education and health most exposed to AI": YTN video passes 20,000 views as comments split over 'gender division'
YTN's Nov. 7 caption news 'Not office work: women-heavy education and health jobs top AI's target list' passed 20,000 views within a day.
Per the video, the Korean Women's Development Institute combined women's job counts by occupation with AI adoption effects. Contrary to forecasts that AI would first replace routine clerical work, education had the highest 'AI risk,' employing 858,000 women, followed by health. It cited the ILO as saying 29% of women's tasks are exposed to generative AI versus 16% for men, and an analysis that women face up to 1.5 times the risk of needing a career change by 2030. It opens with the 1940s AT&T switchboard automation that wiped out 75,000 women's jobs.
Comments split: "Fine for education and clerical work, but AI in health?" got 18 likes, a comment calling it ultimately women's support policy got 76, and "women can do the hard and dangerous work too" got 56. Others objected to dividing by gender.
- ▶AI 1순위 타깃 사무직 아냐... 교육·보건 '여초 직군 초비상' [자막뉴스]YTN · 20K views
Career"I thought liberal arts was over": Samprotv interview on why AI firms seek philosophers hits 58,000 views
Samprotv's Nov. 8 interview with Korea University philosophy professor Seong Chang-won, 'I thought liberal arts was over… why AI companies look for philosophers,' passed 58,000 views.
The description promotes 'philosophical thinking everyone needs in the AI era' and the professor's book. Comments turned to employment: "If philosophy grads get jobs, there's no reason to study it, because philosophy is romance" got 4 likes. One commenter said philosophy graduates mostly go to Coupang fulfillment, tutoring and services, sarcastically calling it enviable. Another said the professor had persuaded the host with philosophy.
- ▶문과는 끝난 줄 알았는데… AI 기업들이 철학자를 찾는 이유 | 성창원 고려대학교 철학과 교수 [주말인터뷰]삼프로TV 3PROTV · 58K views

Feature · Opus 5.5

Is Opus 5.5 really dumber? Complaints pour in, trackers say 'no change,' and a hidden setting sits in between

"The model name didn't change, but the way it does reasoning took a sharp turn." That post on X on Nov. 7 drew about 2,400 likes and reignited the 'Opus 5.5 nerf' debate. Anthropic has never admitted to a downgrade. We checked the complaint numbers, outside trackers, and what the official docs say about default settings.

The post that lit the fuse: 1,400 sessions analyzed by AI
Shane Melaugh posted on Nov. 7 that he had Gemini 4 Argon analyze 1,400 Claude Code sessions and "on November 3, something definitely happened." His claim is that the model name stayed the same but its reasoning changed sharply, and that how much reasoning Opus 5.5 does at "high" dropped. The post passed 380,000 views. The analysis was done by an AI, and its method and sample have not been independently verified.
Three hours vs. minutes: high vs. xhigh
In the same thread Melaugh said that on a real bug, high jumped into action, got confused and failed to find a fix after three hours, while xhigh paused to think and fixed it in minutes. He argued the "use high for everything" rule no longer applies. Developer Theo replied that he told a bot to check the claim against his ClickHouse logs (about 830 likes).
User complaints: plans, limits, 57 mistakes
AIMind said the $20 plan was enough in September but even $100 runs short now, posting two videos from the same prompt (about 2,800 likes). @maxt3chno said limits drain faster because it loops wrong decisions and ignores notes (about 1,500), and greco_dev said he never filled a weekly session at launch but now does in three days. voidfreud claimed 57 mistakes over 3 hours 10 minutes.
The most concrete evidence: one user's 6,638 answers
In GitHub issue #105112, one user said they scored 6,638 of their own answers. Before Oct. 19, 5.2% of 5,738 needed fixing; after, 25.4% of 900. Tool calls and thinking tokens did not fall, and raising effort did not help, they said. They narrowed the shift to the evening of Oct. 18 to afternoon of Oct. 19 (Pacific). No Anthropic employee replied. Issue #105477, posted Oct. 27, reports a similar decline since the week of Oct. 19.
Outside tracker 1: LiveNerf says 'no change'
The open-source LiveNerf has run a frozen 78-question panel daily with a pinned Claude Code CLI since Opus 5.5's Sept. 22 launch. As of Nov. 7 (day 45), all three 10-day windows read 'no change,' the last at +0.028 ± 0.023 versus launch week. But the smallest shift it can detect per window is about 6.6 points, and in testing it could not distinguish a swapped-in Opus 5 at 99% confidence. The effect of medium effort shows up more clearly in output tokens (−26%) than in accuracy.
Outside tracker 2: BridgeBench 'no nerfs detected'
Per i10x, BridgeBench's NerfBench put every Opus 5.5 run between 94.2% and 105.8% of its launch score, reporting "No nerfs detected" on Oct. 16. The same roundup says about 2,700 related posts appeared on X in 24 hours on Nov. 4-6, and that an Oct. 24 post by a former Claude Code team member asking if Opus 5.5 got dumber passed 970,000 views.
The hidden setting: default effort dropped a notch
Anthropic's docs say "Claude Opus 5.5 and Claude Haiku 5.5 default to medium." Opus models before Opus 5 defaulted to high. That means requests that don't set effort run a step lower than on Opus 5. Adaptive thinking cannot be turned off. Anthropic's status page shows one Oct. 24 notice of elevated errors for Opus 5.5 and no quality-degradation notice.
Competing theories and the April precedent
Some users claim Anthropic quantized the model, while others in the same thread argued reasoning compute was reduced; neither side offered evidence. In its April 23 postmortem, Anthropic said earlier Claude Code quality complaints came from an effort change from high to medium, a cache bug and a system prompt, not the model itself. Anthropic has given no public statement on the current complaints.
What you can do
Per the docs, you can set effort to high or xhigh explicitly and compare the same task. Check your Claude Code version, and average several runs of the same prompt rather than trusting one. Reproducible cases can be reported with the /bug command in Claude Code.

AI & Markets

4 stories

Not investment advice. You are responsible for your own investment decisions.

Big TechNvidia board opens probe into CEO Jensen Huang: "checking whether he was targeted by Chinese intelligence"
Nvidia's board has opened an investigation into CEO Jensen Huang, The New York Times reported on Nov. 8, after he acknowledged to the board an improper relationship with a business executive U.S. intelligence determined to be a Chinese spy.
Per Bloomberg, the board is examining whether Huang was targeted by a foreign espionage plot, and he remains CEO for now. On Oct. 29 the Times had reported that U.S. intelligence concluded an executive at a Singapore company passed information to China; Nvidia said at the time that "Jensen talks with thousands of customer executives every year," and China called the claims groundless speculation. That report said officials believed Huang did not know of her activities.
Markets reacted over the weekend: a post said Nvidia fell about 3% in perpetual futures that trade on weekends (about 700 likes). The regular session opens Monday. SK hynix and Samsung Electronics, which supply HBM to Nvidia, are major counterparties. The scope and timeline of the probe are not known.
Further reading
SemiconductorsTSMC denies 3D DRAM move a day later: "core business is foundry," while Samsung and SK hynix watch
After reports that TSMC started planning 3D DRAM production at Nvidia's request, TSMC said its core business is chip contract manufacturing.
Per IT Chosun on Nov. 8, The Information reported on Nov. 6 that TSMC is weighing building a plant or buying one to convert to memory, led by advanced-packaging VP Hsu Kuo-chin, a former Micron Taiwan executive. A TSMC spokesperson told Tom's Hardware, "We always focus on our core business of contract manufacturing."
Unlike HBM, which stacks chips, 3D DRAM stacks memory cells vertically. Logic and DRAM processes differ fundamentally, so securing IP and mass-production experience quickly would be hard, and Tom's Hardware said TSMC's entry would not ease the memory shortage for at least five years. Samsung, SK hynix and Micron are all developing 3D DRAM.
Further reading
- IT Chosun: Thought TSMC was coming for Samsung and SK hynix? TSMC draws a line on 3D DRAM entryIT조선 ↗
- Special Economy: TSMC weighs entering 3D DRAM market with Nvidia스페셜경제 ↗
US MarketsBipartisan Senate bill would fast-track AI data center power infrastructure, colliding with post-midterm backlash
A bipartisan Senate bill to speed permits for AI data centers and power infrastructure could get a vote as early as next week, Investing.com said, citing The Verge (about 2,300 likes).
Reuters reported on Nov. 4 that Democratic wins in governor races in Texas, Georgia and elsewhere raised uncertainty for large data center projects. Texas winner Gina Hinojosa pledged to pause new data center approvals until community-protection laws are in place. Polymarket said restrictions won in 40 of 41 local races, without giving a source.
In Korea, nuclear instrumentation firm Woori Technology fell 3.82% on Nov. 8; CBC News said it was not confirmed that U.S. regulation directly drove the drop. Data center power demand depends on permitting speed, touching transformers, generation equipment and nuclear. The bill text and vote schedule are not final.
Further reading
Timeline signalAlibaba's Qwen 4 27B rumored for 3 p.m. Beijing time Nov. 9; open-model rumor stirs developers
A post saying Alibaba's Qwen 4 27B open model will be released on Nov. 9 at 3 p.m. Beijing time drew about 1,500 likes on X. There is no official confirmation.
K-AI (@OGKAIM) posted the schedule on Nov. 8. A reply calculated that a 27B model quantized to 4-bit needs about 14 GB of VRAM, fitting a single consumer GPU. Another developer said Qwen-4-Flash-Early (codename hawk) was added to GameFeel Bench (about 850 likes).
Another user argued that the open-model combo "Hermes + DeepSeek V4.1 Flash" beats "Claude Code + Fable 5" (about 880 likes); it is a personal opinion with no figures. Whether it ships will be settled this afternoon.
Further reading

All AI & Markets news →

Other AI news

4 stories

Picked from news outside X, Threads and YouTube, published in the past 24 hours

EverydayChatGPT tops 18.76 million monthly users in Korea for the first time; Claude up 15x in a year to No. 2
ChatGPT's mobile app reached 18.76 million monthly active users in Korea in October, passing 18 million for the first time.
Citing Mobile Index data on Nov. 8, IT Chosun and Newneek reported 18,762,276 users, up 43.8% from 13,048,025 a year earlier and 4.3% from September. Total time spent was about 50.98 million hours, more than double a year ago (+113.9%), and average monthly time per user rose from 110 to 163 minutes. New installs fell 16.2% to 791,000.
Claude's app had 2,095,157 monthly users, about 15 times last year's 139,131, and climbed from 7th to 2nd among Korean AI apps in a year. It is still about nine times smaller than ChatGPT. Analysts read the data as a shift from drawing new users to deeper use by existing ones.
Further reading
- IT Chosun: “1 in 3 Koreans use it”: ChatGPT passes 18 million Korean users for the first timeIT조선 ↗
- Newneek: ChatGPT hits 18.76 million Korean users; usage time doubles in a year뉴닉 ↗
EducationNYC mayor Mamdani: "Google using monopoly power to force AI on schools" as it blocks search at schools that restrict AI
New York City Mayor Zohran Mamdani criticized Google for blocking its search engine at schools that restrict its AI features, calling it a use of monopoly power.
Per Futurism on Nov. 8, Mamdani said "No corporation should have the power to dictate how we teach our children." The city announced on Sept. 2 a one-year pause on student generative AI through 8th grade, reported to cover about 40 tools and roughly 600,000 students.
Zephyr Teachout wrote on X that Google is forcing the city to break its own rules and that the city should kick out Google search and refer the case to prosecutors (about 1,700 likes). Google did not comment. Reports differ on the details of Google's conditions.
Further reading
- Futurism: Mamdani Blasts Google for Using Monopoly to Bully Schools That Ban AIFuturism ↗
- X (Zephyr Teachout): NYC has a moratorium on student facing AI for K through 8th gradeX (Zephyr Teachout) ↗
USBloomington, Indiana "superintelligent AI" ordinance is narrower than the 'AI ban' headline
A Polymarket post saying Bloomington, Indiana "bans AI use within city limits" drew about 3,300 likes and 620,000 views after the city council acted on an ordinance against superintelligent AI.
A reply summarized that the ordinance's reasoning and enforcement are more specific than the headline: using or distributing AI smarter than a human in the city is illegal. Per the draft (Ordinance 2026-29), it would apply to services run outside the city and be enforced through civil fines and court orders, with up to $2,500 a day mentioned. Critics say it is unclear who decides what counts as superintelligence. In June the council unanimously adopted a non-binding resolution opposing AGI research without safeguards.
Polymarket also posted that U.S. voters chose to restrict or ban new AI data centers in 40 of 41 local ballot races (about 7,300 likes), without citing what was counted.
Further reading
- X (Polymarket): BREAKING: Bloomington, Indiana bans AI use within city limitsX (Polymarket) ↗
- X (Polymarket): Americans vote to restrict or ban new AI data centers in 40 of 41 local ballot racesX (Polymarket) ↗
CopyrightEminem's publisher and Elvis estate sue Suno: "269 songs, $41 million, voice replicas"
AI music service Suno was hit with copyright suits by Eminem's publisher, Eight Mile Style, and by Elvis Presley's rights holders.
Per DancehallMag on Nov. 8, Eight Mile Style, which controls 269 Eminem compositions, sued in San Francisco federal court seeking about $41 million: $150,000 per work ($40.35 million) plus $672,000 over alleged removal of copyright information. The complaint alleges Suno trained on Eminem material and let users generate similar music and lyrics. The outlet noted this is what is being sought, not what a court awarded.
Per The Verge, Elvis Presley Enterprises and affiliated rights holders filed their own case in the Northern District of California, which also alleges the platform enabled unauthorized replicas of his voice. Suno's response was not reported.
Further reading
- DancehallMag: Eminem’s Publisher Sues AI Music Generator Suno for $41 MillionDancehallMag ↗
- The Verge: Elvis and Eminem rights holders are suing SunoThe Verge ↗

View all 70 collected posts

* Summaries and publishing on this page are done by an automated agent. Check the facts against each story's original links.

Today's stories

Get AI News Daily in your inbox!

Trending on YouTube

AI bubble"AI came for my job, but left with my pension": 'The AI Panic Has Started' video passes 330,000 views

Jobs"Education and health most exposed to AI": YTN video passes 20,000 views as comments split over 'gender division'

Career"I thought liberal arts was over": Samprotv interview on why AI firms seek philosophers hits 58,000 views

Is Opus 5.5 really dumber? Complaints pour in, trackers say 'no change,' and a hidden setting sits in between

The post that lit the fuse: 1,400 sessions analyzed by AI

Three hours vs. minutes: high vs. xhigh

User complaints: plans, limits, 57 mistakes

The most concrete evidence: one user's 6,638 answers

Outside tracker 1: LiveNerf says 'no change'

Outside tracker 2: BridgeBench 'no nerfs detected'

The hidden setting: default effort dropped a notch

Competing theories and the April precedent

What you can do

AI & Markets

Big TechNvidia board opens probe into CEO Jensen Huang: "checking whether he was targeted by Chinese intelligence"

SemiconductorsTSMC denies 3D DRAM move a day later: "core business is foundry," while Samsung and SK hynix watch

US MarketsBipartisan Senate bill would fast-track AI data center power infrastructure, colliding with post-midterm backlash

Timeline signalAlibaba's Qwen 4 27B rumored for 3 p.m. Beijing time Nov. 9; open-model rumor stirs developers

Other AI news

EverydayChatGPT tops 18.76 million monthly users in Korea for the first time; Claude up 15x in a year to No. 2

EducationNYC mayor Mamdani: "Google using monopoly power to force AI on schools" as it blocks search at schools that restrict AI

USBloomington, Indiana "superintelligent AI" ordinance is narrower than the 'AI ban' headline

CopyrightEminem's publisher and Elvis estate sue Suno: "269 songs, $41 million, voice replicas"