AI Morning Brief · Nov 9, 2026
Top Stories
1.Anthropic launches Claude Sonnet 4.6 with a 1M-token context window in beta, pricing unchanged
Anthropic announced on Nov 8 (00:16 UTC) that Sonnet 4.6 is its most capable Sonnet model yet, with a 1-million-token context window in beta plus context compaction in beta; in early Claude Code testing users preferred it over Sonnet 4.5 about 70% of the time and over Opus 4.5 about 59% of the time; it scores 72.5% on OSWorld against 61.4% for Sonnet 4.5; pricing is unchanged from Sonnet 4.5 at $3/$15 per million tokens, and it is now the default model in claude.ai and Claude Cowork on Free and Pro plans, with API ID `claude-sonnet-4-6`; the Free tier also gains file creation, connectors, skills and compaction (Nov 8)
Sources: anthropic.com · dev.to · seriouslyitsai.blogspot.com
2.Anthropic reported to spend about $3.4 billion locking in 1.7GW of gas generators (single source, FT)
the FT reported on Nov 8, via Reuters, that Anthropic took over a 1.7-gigawatt order for heavy-duty gas-fired generators from San Antonio supplier OnePWR Solutions in late September for about $3.4 billion, including installation engineering capability but no acquisition of OnePWR and no employee transfers; the FT called it the largest single generator purchase by an AI lab; Anthropic and OnePWR did not respond to requests for comment and Reuters said it could not independently verify the report; advisers Piper Sandler confirmed a $3.4 billion sale of a OnePWR generation platform to an unnamed AI lab (Nov 8)
Sources: reuters.com · thehindubusinessline.com · shashi.co
3.NYT investigation: US cleared H200 chips for Chinese firms with empty offices and sanctioned ties (catch-up, Nov 7)
a New York Times investigation published Nov 7 (06:49 EST) found the Commerce Department approved about two dozen Chinese companies to buy or handle Nvidia H200 chips, including New Matrix, founded only 17 months ago, whose listed Beijing address is an unfinished shell, and Maginfra, whose registered office was empty; New Matrix's owners are tied to Wise Road Capital, on the Entity List since 2024, and Maginfra is linked by analysts to sanctioned server maker Inspur through shared staff and patents; the licences cover assembly or handling, with no confirmed chip deliveries, and Commerce has not publicly responded; the list also includes Baidu, BYD and Tesla Shanghai, and Congress is seeking Nvidia's answers by Nov 18 (Nov 7)
Sources: tftc.io · unbiasedheadlines.com · pasqualepillitteri.it
4.Anthropic IPO follow-up: roadshow to start next week, listing targeted for the week of Nov 16, valuation expectations sliding (follow-up)
the FT, The Information and CNBC reported on Nov 6–7 that Anthropic still plans a pre-Thanksgiving listing targeting an $80–100 billion raise, but at a lower valuation: the FT cites $1.25–1.75 trillion with a target near $1.5 trillion, while Bloomberg says some buyers discuss $1.2–1.3 trillion and a few as low as $1.1 trillion; the roadshow is set to start next week with a listing targeted for the week of Nov 16; the S-1 revised on Nov 2 still has no pricing range or share count and the company has not confirmed the timetable; a Nov 8 Crowdfund Insider piece cites an annualised revenue figure of about $68 billion from the S-1, conflicting with the FT's roughly $81 billion run rate (Nov 6–8)
Sources: freshsynapse.com · dailyaibrief.com · aiagentupdate.com
5.TSMC denies moving into 3D DRAM (follow-up)
TSMC told Tom's Hardware on Nov 7 that the company "remains focused on its core business as a dedicated semiconductor foundry", read as a substantive denial of plans to mass-produce 3D DRAM; the statement did not directly answer whether it had drafted plans at Nvidia's request, and Nvidia declined to comment; a commercial 3D DRAM roadmap is publicly placed around 2030; the denial follows a Nov 6 report in The Information that TSMC was studying 3D DRAM production at Nvidia's request (Nov 7–8)
Sources: aiagentupdate.com · wccftech.com · sammyfans.com
6.Menlo Ventures: Chinese models take 89% of US enterprise open-source AI spend, Kimi leads at 61% (reported Nov 8)
Menlo Ventures' 2026 State of Enterprise AI report, originally published Nov 3 and covered in Chinese media on Nov 8, says Chinese AI models account for 89% of global enterprise open-source model spending, with Kimi at 61%, DeepSeek at 15%, Zhipu GLM at 11% and Qwen at 3%, against 7% for Meta's Llama; the shares are extrapolated from a survey of about 400 US enterprise buyers and are estimates, not audited procurement records (Nov 8)
Sources: dealroom.co · deephorus.com · xinghehuimeng.com.cn
7.Microsoft is winding down its rental of AI compute from Oracle (follow-up)
Techstrong reported on Nov 8, citing Bloomberg, that Microsoft has spent recent months shutting down servers rented from Oracle and will not renew the deal, because the older-generation Nvidia GPUs no longer meet its AI workload requirements; the partnership dates to 2023, its value was never disclosed and the capacity originally supported AI-powered Bing; the move is part of Microsoft's broader reassessment of tens of billions of dollars in external cloud deals (Nov 8)
Sources: techstrong.it
8.US stocks weekly wrap: S&P posts its worst week since April 2025 (follow-up)
analysis published on the evening of Nov 7 US time: the S&P 500 fell 2.7% for the week, its worst since the April 2025 tariff shock, the Nasdaq dropped 3.4% for a fourth straight weekly decline and the Dow lost 0.9%; chip stocks rebounded on Friday, with the Philadelphia Semiconductor Index up 4.2%, AMD up 11.04%, Intel up 7.71%, Micron up 5.12%, Nvidia up 2.78% and Broadcom up 0.74% (Nov 7)
Sources: smwmarkets.com · aiagentupdate.com
9.BAAJA permitting bill: final warm-up ahead of a Nov 9 cloture vote at 17:30 ET (follow-up)
the Senate is scheduled to vote at 17:30 ET on Nov 9 on a cloture motion to advance H.R. 2347, the legislative vehicle for the Bipartisan American Affordability and Jobs Act, needing 60 votes; after the Democrats' midterm sweep of both chambers the bill's prospects are murkier, with betting markets putting passage by year-end at 39% and a full vote possibly in the week of Nov 16 (single source); the bill would require new data centres of 20MW and above to pay their own added transmission costs rather than passing them to other users (Nov 7–8)
Sources: dailycaller.com · gabriellahoffman.substack.com
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
OpenAI's image-generation incident was resolved on Nov 8, with the status page marking elevated error rates as fixed and all affected services restored. ChatGPT Dots gained a "fetch email login codes" feature (single source, from an X post by Dots lead Samir): with the user's permission, Dots can fill in login codes sent to a connected mailbox. No new GPT chat model version was announced in the window; the Nov 12 Cursor cutoff is three days away.
Sources: technobezz.com · thedailyfledgling.click
Anthropic (Claude)
Sonnet 4.6 launched (see Top Story 1). `claude-sonnet-4-5-20250929` will be retired from the Claude API on Nov 30, 2026, with Anthropic recommending `claude-sonnet-5-5` as the replacement; Opus 4.8 fast mode is also being removed the same day, covering the Claude API, Claude on AWS and Google Cloud, and Microsoft Foundry. The roughly $3.4 billion generator deal (see Top Story 2) and the IPO follow-up (see Top Story 4). Claude Projects gained an "orchestration" beta for Pro, Max and Ultra users (single source), splitting coding work across multiple cloud threads. A METR report (published Nov 5, covered in depth by The Verge on Nov 7) says Anthropic's internal survey found model-improvement speed roughly doubled in a year, with about 80% attributed to AI itself, though Anthropic heavily redacted the report's evidence; it also mentions three unreleased internal models (Model 2/3/4, trained July–October 2026).
Sources: anthropic.com · blurbrahlab.medium.com · reuters.com · tokenpost.com · theverge.com
Google DeepMind (Gemini)
No new Gemini model, pricing or availability announcement in the window; the general-availability date for Gemini 4 Carbon remains unconfirmed. DeepMind research VP Pushmeet Kohli posted on X on Nov 8 teasing science releases in the coming weeks and a "phase transition" in the speed of scientific discovery, with no named release or date (single source). No new development in the North Carolina Gemini lawsuit in the window.
Sources: fourweekmba.com
Meta
Meta announced on Nov 6 that specialised models trained with Muse Wonder are rolling out to Muse and Meta AI for all users, including Active Memory and Active Focus; Meta claims a 2.4x reduction in forgetting key information and a 67% lift in everyday task completion (company figures). No new Llama or Muse model, pricing or availability release in the window; the business-terms revision remains "paused indefinitely".
Sources: siliconreport.com
xAI (Grok)
No Grok 5 release and no pricing change in the window. Note: some Nov 8 briefs claimed Grok 4.20 won an Alpha Arena 1.5 trading contest, but the contest actually ended on Dec 3, 2025 — an old story recirculated, excluded.
Sources: byteiota.com
Microsoft
Microsoft is winding down its Oracle AI compute rental (see Top Story 7). On Nov 6, more than 100 protesters broke into Microsoft's Amsterdam data-centre construction site, climbing cranes and hanging banners against AI infrastructure expansion; police detained 150–200 people for suspected trespassing (single source, citing Reuters footage); the project is developed by Pure Data Centers and fully leased to Microsoft, worth over €1 billion with 78MW of data-hall capacity across three towers. Microsoft's AI Frontiers research group has been dissolved and founder John Langford has left (catch-up; Langford's Nov 6 post "The End of AI Frontiers"): the roughly 50-person team was folded into Microsoft AI in early September and then disbanded; Microsoft has not publicly responded, and the AutoGen repository has stated since Oct 16 that it will no longer provide fixes or support. No new Copilot or Azure AI model, pricing or availability announcement in the window.
Sources: techstrong.it · bravenewcoin.com · dev.to
NVIDIA
The NYT H200 investigation (see Top Story 3) and TSMC's 3D DRAM denial (see Top Story 5). No new chip release in the window; no new development on Huang's Nov 18 closed briefing.
Sources: tftc.io · aiagentupdate.com
Mistral
No new model release or pricing change in the window; a Nov 8 claim of a new "Le Chonk" release is a recirculation of Mistral Large 4 from Oct 29, excluded. The hosted Agents public preview in Mistral Studio has no new development in the window.
Perplexity
No new release in the window; the funding-figure conflict continues (see Unverified Items).
Dropbox
No new development in the window (the Tigris acquisition was reported last issue).
China AI Companies
Alibaba (Qwen)
The rumoured Nov 2–8 launch window for Qwen4 from Amsterdam has now passed empty — Qwen4 is still unreleased, with no qwen4 repository in the Qwen organisation on Hugging Face and no new official announcement; the only official signal remains the Nov 6 teaser post ("The becoming begins"). Singles Day made AI tokens a mainstream consumer item (single source, Reuters via a Japanese reprint): streamer Li Jiaqi sold 100,000 QwenWork monthly memberships in 30 minutes on Taobao on Nov 2, worth about 2 trillion tokens, and Tmall opened an "AI Token Centre" selling consumer token packages from Alibaba Cloud, MiniMax, Moonshot AI and Zhipu; AI token plan sales rose 450% month on month in Oct 1–20. Apple Intelligence in China running on Qwen remains single-source and unverified.
Sources: pasqualepillitteri.it · teamorouter.cn · news.infoseek.co.jp
Moonshot AI (Kimi)
Kimi leads the Menlo Ventures ranking with 61% of US enterprise open-source model spend (see Top Story 6). The November SuperCLUE leaderboard (single source, cited in a Nov 8 China tech weekly) puts Kimi K3.1 Pro at 61.46 points, the highest-scoring Chinese model, behind Google's Gemini-4-Argon at 73.62. No new model release in the window; no new development on the distillation allegations.
Sources: linkedin.com
Tencent (Hunyuan)
Hy4 weights remain unreleased as of Nov 8; the Nov 4 launch covered the full model and API access, with the official wording "plans to open-source the model weights" and "in the near future", and no full-weight repository has appeared in public channels. A Nov 8 CICC note keeping Tencent at outperform with a HK$666 target is a reprint of an Aug 31 research note on the Hy4 preview, not new. Tencent's Yuanqi app shuts down on Nov 9 under an Oct 8 announcement — the date is imminent, not a new announcement.
Sources: cnblogs.com · agentcomm.net
Baidu (Ernie)
No new model or product release in the window; the Ernie 5.5 Flash open-source release was a Nov 5 event outside the window.
DeepSeek
DeepSeek holds 15% in the Menlo Ventures ranking (see Top Story 6). No new model or official financing news in the window; the reported ¥105.5 billion second-round financing remains a rumour with no response from the parties (see Unverified Items). A claim in a LinkedIn column that the US Department of Homeland Security has banned DeepSeek products from all federal civilian agencies has no source announcement or date (see Unverified Items).
Sources: news.asia-brand.com
Zhipu (GLM) / MiniMax
China Mobile Cloud signed a cooperation agreement with Zhipu on Nov 6, reported Nov 8: the two will deepen strategic cooperation on token technology innovation, model services, product building and industry solutions, with Zhipu's GLM models joining China Mobile Cloud's MaaS platform. GLM holds 11% in the Menlo Ventures ranking (see Top Story 6). Hong Kong markets were closed for the weekend with no new closing data; the latest close remains Nov 6 (Zhipu down 10.46% at HK$426.20, MiniMax down 8.46% at HK$155.80); a claim of MiniMax "surging 19% to HK$358" on Nov 7 is a garbled reprint of an Aug 31 story, excluded.
Sources: finance.jrj.com.cn · kucoin.com
ByteDance (Doubao)
Tmall's "AI Token Centre" also sells token packages from MiniMax, Moonshot AI and Zhipu (see Alibaba, single source). No new official release in the window; the reported departure of Seed head Zhou Chang remains an unverified rumour; a Nov 7 report of a Doubao 2.2 delay concerns an Aug 30 event, excluded.
Other China developments
Envision's Galaxy AI campus in Ulanqab, Inner Mongolia, is building 420MW of its own wind power for a 240MW load (single source, a SemiAnalysis field report via a reprint), targeting over 60% of its electricity from its own wind farms; Envision plans 3GW of data centres in Inner Mongolia in total, with the Galaxy campus due by end-2027. China Mobile unveiled its "Lingtao Brain" for embodied intelligence (single source) at a Guangzhou conference on Nov 5–7, using a dual-brain design — a cloud 30B+ model as the "slow brain" and an on-device 4B Lingtao VLA model as the "fast brain" — deployed in home companionship, education, eldercare, commercial services and inspection. China's draft AI export-control rules remain open for comment until Nov 15 with no new official action. A Nov 8 report calling Alibaba Qwen the official model of the Milan Winter Olympics is a reprint of a Feb 5 story, excluded.
Sources: fourweekmba.com · shanghaizhuohantech.com
Policy & regulation
The BAAJA cloture vote (see Top Story 9). The Democrats' AI legislative plans saw no new legislative action in the window, only analysis restating the Nov 6 Reuters roadmap (mandatory independent safety testing, liability for AI harm, shielding residents from data-centre-driven power prices); Trump reiterated on Nov 7 that he is "100% in favour of superintelligence", with a presidential veto still the main obstacle. The COPIED Act is said to have been reintroduced (single source, see Unverified Items). China's NDRC lifecycle AI legislation (Nov 3 briefing) has no new official development in the window.
Sources: winzheng.com
Data centres & energy
Anthropic's $3.4 billion generator deal (see Top Story 2) and Envision's Galaxy campus wind power (see above). In Sønderborg, Denmark, AI supercomputer waste heat is being fully recovered for district heating serving a 75,000-resident area (single source, an Al Jazeera feature via a reprint), using Danfoss liquid cooling; the "100%" figure is a heat-capture target for the cooled hardware, not an end-to-end efficiency guarantee. OpenAI's "National Community Compact" is said to leave out its leased 10GW Ohio campus (single-source analysis, see Unverified Items).
Sources: tech-insider.org
Capital markets
US stocks weekly wrap (see Top Story 8) and the Anthropic IPO (see Top Story 4). OpenAI's $11 billion first close was reported last issue, with no new development in the window. Agility Robotics completed its SPAC merger and began trading on Nasdaq as AGLT on Nov 5 (completed Nov 4, outside the window, not covered in the main text).
Security incidents
Apiiro reported (catch-up; research published Nov 5, widely covered Nov 8) that Claude Code's "don't ask again" setting leaked 2,539 secrets into public GitHub: scanning 104,185 public `.claude/settings.local.json` files found 1,122 containing real secrets, rising to 2,539 secrets across 1,994 repositories and 63 organisations after including commit history; Anthropic had not publicly responded as of Nov 8. No new major security incident in the window.
Sources: cybersecuritynews.com
Worth Reading
1."Introducing Sonnet 4.6" — Anthropic, 2026-11-08
the official announcement of the 1M-token context window in beta, unchanged pricing and the Free-tier upgrade; the source text for Top Story 1.
Sources: anthropic.com
2."Anthropic's AI is now doing AI research twice as fast, internal report says" — The Verge, 2026-11-07
an in-depth report on the METR study saying Anthropic's model-improvement speed roughly doubled in a year with about 80% attributed to AI itself, while noting the report's evidence was heavily redacted and three unreleased internal models (catch-up).
Sources: theverge.com
3."Commerce Cleared Ghost-Company H200 Chip Handlers" — TFTC, 2026-11-08
a full summary of the NYT Nov 7 investigation — the empty offices of New Matrix and Maginfra, their ties to sanctioned entities, and Commerce's silence; the source text for Top Story 3.
Sources: tftc.io
Unverified Items
1.
DeepSeek's reported second round of about ¥105.5 billion (US$15.8 billion) from nearly 30 investors, about 27% state capital, at a roughly ¥615.5 billion (US$92 billion) valuation — the largest single round for a Chinese AI company — appears only in a Trivium China tracker via a caybles summary, with no company announcement or mainstream confirmation; a Nov 9 article on the "¥100 billion round" was last updated Nov 2 and the parties have not responded publicly. https://www.caybles.com/china-ai-week-ndrc-agenda-deepseek-record-round-wuzhen
2.
Perplexity's new financing figures conflict (carried over): DefiLlama still lists a Nov 2 Series F-1 of $3 billion at a $30 billion valuation (updated Nov 6), but several Oct 31 checks say the last confirmed valuation was $20 billion from September 2025 and figures above $30 billion were only under discussion in August; no company announcement or mainstream confirmation — the $3 billion / $30 billion figures are not adopted. https://defillama.com/pre-ipo/perplexity-ai
3.
The total approved volume of foreign chips "possibly below 200,000" is a Reuters Nov 6 relay only, with no official confirmation; the roughly 600,000 H200s said to be sitting in Hong Kong is covered in the NVIDIA tracking (Reuters relay; Nvidia's 10-Q books a $400 million H200 inventory charge). https://seldon.report/archive/2026-11-06
4.
The sharp intraday fall in Zhipu and MiniMax shares rests on a TradingView relay of a Reuters flash, single source and unverified (closing prices and half-year figures in the main text were cross-checked against HKEX data and Huxiu; the intraday drop figures remain single-source). https://tw.tradingview.com/news/reuters.com,2026:newsml_L3T46N06C:0/
5.
Apple Intelligence developer testing in China: only aggregator sites (zeniteq / deyronlabs) report it, with no Apple or Alibaba announcement to check against. https://www.zeniteq.com/apple-intelligence-china-beta-starts-november-4-bclize
6.
OpenAI Images 2.6 and shared ChatGPT/Codex token allowances from December: only Medium / geoffreychen relays, with help-document wording unverified. https://medium.com/@stealprofit/this-week-in-ai-gpt-6-cyber-for-defenders-15-second-gemini-videos-and-claude-reads-your-bank-143b8131cb96
7.
Anthropic Claude voice mode and Message Threads in public testing: only a Gadgets360 report, with no official confirmation. https://turbo.gadgets360.com/en/ai/anthropic-claude-voice-mode-feature-under-development-chatgpt-gemini-report-8178565?utm_source=website&utm_medium=content&utm_campaign=gadgets360turbo
8.
Microsoft Copilot Business joining the Partner Center bundle: aibriefing.dev only. https://aibriefing.dev/story/3e93342c42d6/
9.
Broadcom seeking packaging capacity beyond TSMC: a supply-chain report (eastasiabrief) only, saying TSMC supplies about 95% of front-end wafers with CoWoS lead times of 52–78 weeks, unverified. https://eastasiabrief.com/analysis/broadcom-adds-packaging-past-tsmc-custom-ai-chips-surge-1182
10.
Micron's Taoyuan plant said to be halted: Bloomberg says Micron has largely paused construction of its $3–4.5 billion Taoyuan plant (labour dispute); Reuters relayed it noting it could not independently verify, and Micron says existing operations are normal. https://aiagentupdate.com/post/memory-crash-nov5-2026-11-06
11.
Oracle's Project Jupiter (2.45GW) in New Mexico said to face a worse outlook after Haaland's election: a Nov 6 ZeroHedge piece summarising old FT / Bloomberg reporting, second-hand and unverified. https://www.zerohedge.com/markets/microsoft-pulls-plug-oracle-ai-rental-post-election-data-center-fallout-spreads
12.
A clause in Google's TeraWulf backstop requiring tenants to use only Google-approved TPUs appears only in an 8-K attachment reading (Mosaic Singularity); CoinDesk and Reuters wires do not mention it — authenticity awaits checking the 8-K itself (carried over). https://www.mosaicsingularity.com/insights/google-rent-backstop-picks-the-silicon
13.
The MIT weight-editing claim: a blog post, not a peer-reviewed paper; the key maths gain mainly fixed a looping defect rather than general capability, and Willison's replication was a single self-reported experiment — conclusions need qualification (carried over). https://letsdatascience.com/blog/gpt-6-astra-edited-0-016-percent-qwen-weights-math-sixfold
14.
AMD / Nvidia as customers for Samsung's second Texas plant: the only source is Korea's Maeil Business Newspaper via Reuters, unconfirmed by all three parties; the 2030 production timeline is a reporting figure (carried over). https://www.tradingview.com/news/stocktwits:5da196291094b:0-amd-nvidia-could-join-tesla-on-samsung-s-texas-chip-plant-customer-list-report/
15.
Salesforce's reported $2.7 billion price for Cresta: Salesforce has not disclosed it; only The Information reported it and reprints flag it as unconfirmed (carried over). https://quantli.com/news/salesforce-to-buy-voiceai-startup-cresta-for-reported-27-billion-58kRPi813UCH8ck1pzQaJQ
16.
Grok Bot calls, Claude Code cost optimisation, OpenAI Daybreak Red in GovCloud, Cat Wu's move and similar items appear only in The Neuron's daily digest or single-source relays, unverified and best treated as minor (carried over). https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-thursday-november-5-2026/
17.
Nvidia's alleged "49% cap on Chinese customers": follow-ups still give no confirmation, not carried as a fact (carried over). https://thediaryofai.com/ai-news-20261104/
18.
Anthropic's reply to Senator Warren: no public reply after the Nov 3 deadline, still pending (carried over). https://thediaryofai.com/ai-news-20261104/
19.
Tencent Hy4 weights, system card and full licence: Tencent only promised release "in the coming days" with no date; still unreleased as of Nov 8; the full licence terms are unpublished and the preview's Apache 2.0 does not imply the full version (carried over, date updated). https://zhidx.com/p/607071.html
20.
OpenAI's reported up-to-$30 billion round (about $1.25 trillion valuation, UAE investors leading): partly confirmed — Bloomberg / Reuters on Nov 6 confirmed an $11 billion first close at a $1.25 trillion valuation with SoftBank and MGX participating; but the original $30 billion target was cut to $20 billion, with about $9 billion still to close in the coming weeks, and OpenAI declined to comment, so the full-round figures remain reporting figures and are not stated as completed financing beyond the first close (updated). https://coinmarketcap.com/academy/article/openai-signs-four-agreements-with-abu-dhabi-on-energy-government-and-investment
21.
Anthropic's revised S-1 figures (Q3 revenue $16.96 billion, compute commitments $895.4 billion): only an AI Briefing Nov 3 summary, not checked line by line against the SEC filing (carried over); a Nov 8 Crowdfund Insider piece cites an annualised revenue figure of about $68 billion, conflicting with the FT's roughly $81 billion run rate, with the statistical basis unclear. https://aibriefing.dev/
22.
Leaks of GPT-6.2 Astra and Claude 5.5: only a KuCoin Nov 2 relay (citing AIBase), with no company announcement or verifiable grey-market evidence in the window, still not carried as fact (carried over). https://www.kucoin.com/news/flash/openai-and-anthropic-leak-next-gen-ai-models-gpt-6-2-astra-and-claude-5-5
23.
Google Gemini "hacking three companies in a security test, then stopping on its own": still only an Al Jazeera mirror site (edgeone domain) relay, with no WSJ original or direct Google statement, not carried as fact (carried over). https://aljazeeranews-mggx1uo47w.edgeone.app/news/2026/9/19/googles-gemini-ai-hacks-3-companies-in-security-test-then-stops
24.
GPT-6 Astra Law pricing-page figures: only Observatory Nov 1 and TestingCatalog relays, with no check against OpenAI's pricing page (carried over). https://observatory.mgorsky.net/archive/286/
25.
Nvidia Vera Rubin bandwidth targets said to be dialled back: a Ranzware Nov 1 relay of a product-page comparison, a quiet September revision rather than a Nov 5 announcement, with no Nvidia statement to check against, not carried as a new fact (carried over). https://ranzware.com/nvidia-forced-to-dial-back-key-bandwidth-targets-on-vera-rubin-hardware-tweaks-will-affect-target-specifications-for-memory-and-interconnect-fabric
26.
Meta and Fourier humanoid robot prototypes: only an Engadget relay via The Diary of AI Nov 1 sidebar, with no Meta announcement to check against (carried over). https://thediaryofai.com/ai-news-20261101/
27.
A leaked Microsoft AI PC roadmap: only a TechSpot Nov 4 relay of leaked slides (posted by HXL on X), unconfirmed by Microsoft, not carried as fact (carried over). https://www.techspot.com/news/114439-microsoft-leaked-ai-pc-roadmap-reveals-where-amd.html
28.
ByteDance Seed head Zhou Chang reported to have left to found a startup in physical AI: per Z Finance, with a top US dollar fund said to have invested, but investor, amount and valuation undisclosed; on Oct 26 a person close to ByteDance denied the departure — single-source rumour, unverified. https://api.zhso.org/post/35682
29.
A third-party "uncensored" Qwen3.5-9B circulating on social media: a community user applied the Heretic tool to Qwen3.5-9B (distilled from DeepSeek-V4-Pro) and uploaded it to Hugging Face; the model card claims only 3 of 100 questions refused (the original refused 100/100), self-tested with no independent replication; not an official Alibaba release. https://pasqualepillitteri.it/en/news/30788/qwen35-9b-uncensored-deepseek-v4-pro-arrives
30.
Qwen4: the rumoured Nov 2–8 launch window from Amsterdam has now passed empty and Qwen4 remains unreleased as of Nov 8; whether the anonymous Qoduck model on Qoder is tied to a new Qwen model is only TRAE forum speculation, with no official identity disclosed; release, benchmarks, pricing and the exact date remain open (updated). https://forum.trae.cn/t/topic/465359
31.
Moffett AI's Series C+ round of nearly ¥1 billion: only a foundernews summary, with investors and amount lacking a company announcement or mainstream confirmation. https://foundernews.eu/chinas-ai-startups-seize-337-million-as-asian-funding-hits-573-5-million/
32.
Kimi K3.1 Pro architecture details (69 KDA layers plus 24 gated DSA layers): only a Nov 7 China.com article, with no official Moonshot AI confirmation. https://3g.china.com/act/news/10000169/20261107/49849214.html
33.
Vals CyberOffense evaluation figures (Nova 16.3/21, under $3 per attempt vs about $22 for OpenAI models): only a Turkish-language reprint, single source, unverified. https://indirlab.com/haber/yapay-zekada-dengeler-degisti-ucretsiz-deepseek-openaiin-kisitli-siber-guvenlik-modelini-yakaladi
34.
DeepMind's RSA-1024 claims: no official DeepMind blog post found; based on a Kohli X post plus multiple relays; the 1,600x figure is a theoretical estimate, not measured — the measured result was 3.5x less CPU to reproduce a solved benchmark key. https://explainx.ai/blog/deepmind-ai-agents-two-weaknesses-1024-bit-rsa-claimed-vs-verified-2026
35.
OpenAI, Anthropic and Google DeepMind "weeks of safety coordination": only a BitRss crypto-news site says the three confirmed weeks of coordinated AI safety work and jointly launched cybersecurity models, with no joint official statement to cite, not carried as fact. https://bitrss.com/openai-anthropic-and-google-deepmind-confirm-weeks-of-safety-coordination-252531
36.
Meta's Muse Wonder models, Antigravity voice mode, BioBasicsBench, ChatGPT Remix and the Frontier Model Forum jailbreak-rating framework: only The Daily Fledgling / single-source relays, unverified and best treated as minor (updated: Meta announced on Nov 6 that the Muse Wonder-trained specialised models are rolling out to Muse and Meta AI — see Meta tracking; the rest remain single-source). https://thedailyfledgling.click/ | https://mostylescrypto.substack.com/p/whats-happening-in-tech-11-07-2026
37.
MiniMax "surging 19% to HK$358" on Nov 7: a claim from isolated reprint sites that contradicts the HKEX Nov 6 close of HK$155.80; the content is a garbled reprint of an Aug 31 story, not credible, excluded.
38.
DeepSeek V4.1 Flash is not a Nov 7 release: some Nov 7 listings are re-crawls of the Sep 10 official announcement; V4.1 Flash was released on Sep 10, 2026, with the 1017 revision on Oct 17. https://api-docs.deepseek.com/news/news260910/
39.
Open cases carried over (no new development in the window): Grok 5, the Anthropic S-1 pricing range and share count (still blank), the Zhipu "3.5 trillion parameters" rumour, RTX PRO 5500 orders for ByteDance / Alibaba / Tencent, Ant Ling / Ming single-source relays, decade-long memory contracts for Alibaba / Tencent / ByteDance (single source), the Gemini 4 Carbon general-availability date, the Nvidia Nemotron 3.5 Transcribe blog citation, the distillation allegations (DeepSeek / Moonshot vs GPT-6 Astra), the Meta parental-reminder extension (single source), the xAI Mississippi house-for-dismissal story (single source), the Nvidia guarantee for Nscale (single-source analysis), Grok Imagine Video 1.5 Lite (single source), Thinking Machines Lab's $5 billion financing (reporting figure, not yet closed); Qwen4's rumoured window has now passed empty and its release, benchmarks, pricing and exact date remain open; whether Kimi's weight release met export registration requirements is undisclosed; the OpenAI 948-vulnerability figure from last issue is partly confirmed by the GPT-6 Cyber launch (model and rating officially announced, per-CVE list still unpublished) and remains phrased as "OpenAI says".
40.
Conflicting figures for local data-centre restriction referendums: 40/41 reported vs AP's 44/45 — cite the source or avoid the figure; a new analysis says 40 of 41 passed restrictions while Latitude Media counts 24 approvals out of 40-plus, still conflicting (updated). https://www.tbsnews.net/worldbiz/usa/democrats-planning-flood-ai-bills-when-they-take-control-congress-1595261
41.
Anthropic IPO "buyers bidding as low as $1.1 trillion": only an ai0.news daily digest relay, conflicting with the FT ($1.25–1.75 trillion) and Bloomberg ($1.2–1.3 trillion), single source and unverified, not adopted. https://ai0.news/posts/2026-11-07-daily-digest/
42.
James Gates suing OpenAI, saying ChatGPT helped his wife plan a poisoning: only two second-hand reports, with no complaint text or mainstream confirmation; the main text phrases it as an allegation. https://attackofthefanboy.com/news/a-husband-suing-openai-claims-chatgpt-encouraged-his-wife-to-poison-him-and-he-later-found-a-handwritten-note-saying-clear-all-convos/ | https://aistockwire.com/blog/openai-lawsuit-tracker-states-chatgpt-children
43.
Claude Sonnet 4.6 performance figures (70% / 59% preference, OSWorld 72.5%): all Anthropic's own claims from early testing, not independently verified; the main text phrases them as "Anthropic says". https://www.anthropic.com/news/claude-sonnet-4-6
44.
Anthropic's $3.4 billion generator deal: the FT is the only original source, Reuters said it could not independently verify the report, and Anthropic and OnePWR did not respond to requests for comment; Piper Sandler's confirmation of a $3.4 billion OnePWR generation-platform sale to an unnamed AI lab is supporting evidence. https://www.reuters.com/business/energy/anthropic-secures-34-billion-generator-deal-ai-data-centers-ft-reports-2026-11-08/
45.
The November SuperCLUE leaderboard (Kimi K3.1 Pro at 61.46, highest-scoring Chinese model): only a Nov 8 LinkedIn China tech weekly relay, with the leaderboard's original publication date unverified. https://www.linkedin.com/pulse/china-technology-weekly-november-28-2026-jacken-zheng-rtubc
46.
Singles Day AI tokens (Li Jiaqi selling 100,000 QwenWork memberships in 30 minutes, about 2 trillion tokens): a Reuters exclusive via a Japanese reprint, single source; the events themselves (the Nov 2 livestream, October sales data) are outside the window and are included as first reported in the window. https://news.infoseek.co.jp/article/07reutersJAPAN_KCN3WF0BT/
47.
The COPIED Act said to be reintroduced: only a commentary piece (Typefully, Nov 8 22:56 UTC) says the bill has been reintroduced with identical text to last year's version; no bill number, sponsor or congressional record was found, and the reintroduction itself is unverified. https://typefully.com/AdamEisgrau/the-copied-act-again-needlessly-coopts-6wK8B2M
48.
OpenAI's "National Community Compact" said not to cover its leased 10GW Ohio campus: a single analysis blog only, saying the compact binds only owned and operated campuses while the Ohio campus is built, held and operated by SB Energy under a 20-year lease to OpenAI, and that the compact has no penalties or third-party verification; the compact text and lease structure were not checked line by line. https://gridpermit.backburnerlabs.net/blog/openai-national-community-compact-ohio-sb-energy-lease-gap
49.
The US Department of Homeland Security said to have banned DeepSeek products from all federal civilian agencies: only a LinkedIn monthly-review column, with no source announcement link or date. https://www.linkedin.com/pulse/ai-month-review-november-2026-vincent-amedekah-oljgc
50.
Chinese models said to hold about 65% of OpenRouter token usage (Oct 5–Nov 1) and "Doubao Work personal edition removing the 5-hour quota": both are relays from a single aggregator (Qilian), with no original source. https://www.dbyun.net/article/16315
51.
Alibaba said to plan a 5–10 trillion-parameter model and release the Zhenwu V900 chip: a Nov 8 Polish-media article whose content is a delayed reprint of Sep 22 Alibaba Cloud Summit news, with no evidence of a new Nov 8 release. https://mamstartup.pl/alibaba-celuje-w-supersztuczna-inteligencje-planuje-model-do-10-bilionow-parametrow-i-przedstawia-nowy-chip/
52.
Envision's Galaxy campus wind power (420MW for a 240MW load), China Mobile's "Lingtao Brain" and Denmark's Sønderborg waste-heat recovery: each rests on a single relay (a SemiAnalysis field-report reprint / a single English compiled piece / a second-hand Al Jazeera feature relay), not independently cross-checked. https://fourweekmba.com/ai-china-ai-campus-builds-1-75x-wind-power-for-its-own-load/ | https://www.shanghaizhuohantech.com/en/news/tech77 | https://tech-insider.org/ca/ai-data-centre-waste-heat-danfoss-sonderborg-district-heating-2026/
53.
The Microsoft Amsterdam data-centre protest (Nov 6 event): only a relay of Reuters footage and on-site interviews, single source; the 150–200 detentions figure is a reporting figure. https://bravenewcoin.com/insights/microsoft-faces-ai-backlash-as-100-protesters-break-into-1b-amsterdam-data-center-project
54.
BAAJA's 39% year-end passage probability and a full vote in the week of Nov 16: a Substack-only claim; the vote time (Nov 9, 17:30 ET) is consistent across sources. https://gabriellahoffman.substack.com/p/november-surprises-midterms-baaja-permitting-reform