AI Morning Brief · Nov 5, 2026
Top Stories
1.Tencent releases full Hy4 on Nov 4
the first Hunyuan model with vision and computer use: a 770-billion-parameter mixture-of-experts system (up from 295 billion in Hy3) that reads fine-grained images, complex charts and mathematical visuals, and operates computers through command-line and graphical interfaces; live in Yuanbao, WorkBuddy, CodeBuddy, ima and WeCom's Dayuan assistant, plus Tencent Cloud TokenHub and third-party APIs including OpenRouter, at unchanged pricing ($0.834 per million input tokens, $2.501 output, $0.042 cached); weights and the full system card are promised only "in the coming days" and the full-release licence is unstated (the August preview was Apache 2.0); vendor-reported hallucination rate fell from 15.76% for the preview to 5.45%, with Terminal-Bench 4.0 at 44.4 against 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol (Nov 4)
Sources: zhidx.com · unicornapi.net · madrobot.blog
2.Moonshot AI publishes Kimi K3.1 Pro weights on Nov 4
meeting its promised date: Hugging Face repository moonshotai/Kimi-K3.1-Pro, 96 safetensors shards totalling about 1.56 TB, a 2.8-trillion-parameter mixture-of-experts model with 104 billion active per token, 896 experts (16 per token), 93 layers (69 KDA plus 24 Gated DSA), a 1,048,576-token context and MXFP4 weights from quantization-aware training; the custom Kimi K3.1 Pro License is MIT-style but requires a separate Moonshot agreement for model-as-a-service businesses above $20 million revenue in 12 months, and on-interface attribution for products above 100 million monthly users or $20 million monthly revenue; API pricing is unchanged at $2 input / $10 output per million tokens (Nov 4)
Sources: mlllm.io · explainx.ai · thediaryofai.com
3.US midterm voters reject data centers almost everywhere on Nov 3 (reported Nov 4)
of about 41 local ballot measures across 10 states, all 35 with posted counts by the morning of Nov 4 went against development, with trackers counting at least 33 restrictions passed and six more leading; the only exception was Harwood, North Dakota, where 75% of 430 voters rejected a review amendment; all 21 Ohio measures passed restrictions, including Trenton's ban at about 80%; NBC exit polling found 56% of voters called AI and data centers a crucial or major factor and 76% opposed a new data center near home; results are unofficial and uncertified, and some measures are advisory only (Nov 3–4)
Sources: usatoday.com · route-fifty.com
4.OpenAI agrees to acquire Pierre on Nov 4
the developer-tools company behind the Trees, Diffs and Highlights libraries already running in the Codex app, plus the Git-based Code Storage service for machine-oriented code storage (vendor specs: 32 TB repositories, 30+ concurrent writes, Git LFS); no price disclosed, closing subject to customary conditions including regulatory approval, the companies staying independent until close; open-source projects stay open source and Code Storage customers can keep using it while its features move into Codex; the deal follows OpenAI's Nov 2 disclosure of Pando, its in-house Git backend (Nov 4)
Sources: unite.ai · siliconnoon.com · blockchain.news
5.Meta acquires Conductor on Nov 4
the startup whose Mac app runs several coding agents (Claude Code, Codex, Cursor, OpenCode) in parallel isolated code copies, with founders Charlie Holtz and Jackson de Campos joining Meta Superintelligence Labs; terms undisclosed; the product winds down on a fixed schedule — new sign-ups and downloads stopped Nov 4, Cloud workspaces read-only from Nov 13, Cloud shutdown and data deletion beginning Nov 20, local Mac app supported until early 2027; the same day OpenAI announced the Pierre deal, making it a double acquisition day for the coding-agent tooling layer (Nov 4)
Sources: techcrunch.com · unite.ai · siliconangle.com
6.Google's school policy, announced Nov 3, draws fire on Nov 4
Gemini, Gemini Notebook, Gemini in Classroom and Search will be off by default for K-12 student accounts until administrators switch them on, taking effect in early 2027 (New York Times account), but because Search and AI are deeply integrated, schools that decline the AI tools also lose Search; Senator Bernie Sanders called it "extortion," Senator Richard Blumenthal said Google cares more about forcing Gemini into schools, former FTC chair Lina Khan called it anticompetitive coercion, and New York mayor Zohran Mamdani objected to corporate control over teaching; Google clarified on X on Nov 3 that schools can switch on Gemini, NotebookLM and Search individually, and that Guided Learning went live globally in Classroom on Nov 3 (Nov 3–4)
Sources: techcrunch.com · medianama.com
7.Anthropic's FlashProve, widely reported Nov 4, cryptographically proves which model wrote a reply
a research prototype from Anthropic's Frontier Red Team with EPFL, MIT, StarkWare and Succinct: providers publish a weight commitment (fingerprint), and each reply carries a zkSNARK proof checkable without a GPU and without exposing weights; tested on Meta's Llama 4 Maverick (400B mixture-of-experts, 4-bit public conversion), proving runs about 110 tokens per second on one H100, proofs measure about 19 MB per 1,024-token request and verify in about 17 seconds on 24 CPU cores, at roughly 70–95x normal serving GPU time; proofs cover only contexts up to 1,024 tokens with greedy decoding of an integer model version, and Anthropic has no deployment plans — adoption depends on labs publishing fingerprints (code on GitHub since Oct 30, widely reported Nov 4)
Sources: interestingengineering.com · nogentech.org · thinkfacility.com
8.Fal raises $1.3 billion in Series F (announced Nov 3, widely reported Nov 4)
Bessemer Venture Partners leads the round for the San Francisco generative-media inference platform, with Nvidia, Sound Ventures, HOF Capital and T. Rowe Price participating, valuing Fal at $16 billion, double its $8 billion April valuation; disclosed together with a previously closed $350 million Series E led by Sequoia, the two rounds total $1.65 billion and cumulative funding exceeds $1.9 billion; Bloomberg's account puts annualized revenue above $800 million earlier this year, and the company says the money funds compute and hiring beyond its current 185 staff after summer demand forced customer rate limits (Nov 3–4)
Sources: tech.eu · webpronews.com
9.ChatGPT gains GIF creation with Images 2.6, reported Nov 4
The Verge reports OpenAI is rolling out animated GIF generation in ChatGPT from text prompts or uploaded photos, with playback-speed controls and frame-by-frame editing and sharing, alongside faster image generation; the rollout covers ChatGPT and Codex on web, desktop and mobile but is gradual and most users cannot access it yet, and OpenAI has published no standalone product announcement; developers get two new API models, GPT-Image-2.6 Flare (speed) and Sunburst (quality); OpenAI's own tally puts weekly image output across ChatGPT and the API above 4 billion, up from 3 billion in September (Nov 4, single outlet report of an unannounced model)
Sources: unite.ai · decrypt.co · xda-developers.com
10.AMD reports record Q3 revenue of $13.38 billion (released Nov 3, reported Nov 4)
up 45% year on year and above roughly $13.09 billion expected, with non-GAAP earnings of $2.09 per share against about $1.95 expected; data-center revenue hit a record $8.55 billion, up 97% and 64% of sales, driven by EPYC processors and first Instinct MI450 accelerators in Helios rack systems, which AMD says Anthropic, Meta, Microsoft, OpenAI and Oracle are already running workloads on; client and gaming revenue fell 8% to $3.7 billion, gaming alone down 40% to $773 million; guidance calls for about $17 billion in Q4 (±3%) with data-center revenue up about 50% sequentially, and CEO Lisa Su expects total revenue growth above 80% in 2027 (Nov 3–4, guidance is a management estimate, not a result)
Sources: zacks.com · techspot.com · invezz.com
Platform & Model Tracking
OpenAI (GPT/ChatGPT)
Pierre acquisition (see Top Story 4); Images 2.6 GIF feature (see Top Story 9). Catch-up from Nov 2 (not covered last issue): Pando, its in-house Git backend, now serves all production traffic for its monorepo, cutting median push latency by nearly 80% and serving almost a petabyte of repository data a day. Security follow-up (announced Nov 2, Reuters investigation Nov 3, covered Nov 4): Cathedral Systems researchers will use OpenAI's models alongside internal teams in the next security sprint, with bug bounties raised more than fourfold for serious bugs and top reports up to $172,000; Cathedral, founded by former DOGE staffers, reached parts of Anthropic's internal systems in under 12 hours in October via third-party business software — Anthropic says only sales, HR and IT systems, not weights or Claude conversations, and disputes Cathedral's disclosure; Reuters reported Nov 3 that employees had warned since before ChatGPT's 2022 launch that sensitive information was too widely accessible inside OpenAI, which disputes any neglect of security; sprint scope and dates are undisclosed. Catch-up from Nov 2 (reported Nov 4): consumer product lead Adam Fry put ChatGPT search-like queries at roughly 20% of Google's search volume (OpenAI's own estimate by its definition of search-like activity, not an independently audited share), with 1.2 billion weekly users and 250 million weekly voice-input users, and said Chat and Work will merge into one experience. Personnel: former Uber executive Jill Hazelbaker appointed chief corporate affairs officer, reporting to President Greg Brockman. No new GPT model version announced in the window.
Sources: unite.ai · theneuron.ai · madrobot.blog
Anthropic (Claude)
FlashProve research prototype (see Top Story 7). Personnel: Claude Code and Cowork product leader Cat Wu announced her departure on Nov 3 (Business Insider, reported Nov 4), her last day the preceding Friday, with no public reason given. Senator Warren's reply remains publicly unanswered in The Diary of AI's Nov 4 follow-up. No new Claude model, pricing or availability announcement in the window; IPO price range and share count remain blank.
Sources: interestingengineering.com · thediaryofai.com
Google DeepMind (Gemini)
K-12 school policy (see Top Story 6). Follow-up: the Washington Post reported Nov 4 that current and former DeepMind employees challenged contractual language appearing to restrict public criticism after departure; Google says it does not enforce post-employment restrictions and has removed the provision from new contracts, while employees want the same assurance for existing agreements. No new Gemini model, pricing or availability announcement in the window; Gemini 4 Carbon's general-release date remains unconfirmed.
Sources: techcrunch.com · theneuron.ai
Meta
Conductor acquisition (see Top Story 5). No new Llama or Muse model, pricing or availability release in the window; the planned up-to-$95 billion equity financing remains a plan; the Fourier humanoid prototype remains single source (unverified).
Sources: techcrunch.com
xAI (Grok)
No Grok 5 release and no pricing change in the window; the Nov 7 default watermark for the Grok API is a previously reported future effective date with no new wording; no new ruling in the NAACP turbine lawsuit.
Microsoft
Frontier Launchpad follow-up Nov 4 (TechMonitor): applications open until Nov 27, first cohort starting early 2027, across AI platforms and fintech paths with Microsoft and Nvidia technical experts and Marketplace sales guidance — an ecosystem program, not a model release. Copilot Autopilot reaches Frontier in November (Microsoft's October Copilot update, revised Nov 3), labelled private preview with tenant availability subject to admin confirmation. A leaked AI PC roadmap (TechSpot Nov 4, citing a leaked slide posted by HXL) placing Ryzen AI 500 and Intel Nova Lake in the mainstream Copilot+ tier is single source and unconfirmed by Microsoft (unverified). No new Copilot or Azure AI model, pricing or availability announcement in the window.
Sources: techmonitor.ai · techspot.com
NVIDIA
No new chip release in the window; no new official development in the Justice Department inquiry into the proposed Hugging Face acquisition. The reported 49% cap on Chinese customers remains unconfirmed in The Diary of AI's Nov 4 follow-up (unverified). Nvidia participated in Fal's Series F (see Top Story 8); Frontier Launchpad with Microsoft (see Microsoft).
Sources: thediaryofai.com
Mistral
No new model release or pricing change in the window; the Large 4 scoring dispute has no new wording in the window (vendor 41.4% vs independent 24.7% and 26.26%, unresolved).
China AI Companies
DeepSeek
No new official status update in the window on the Nov 3 afternoon outage — the "fixed but still degraded" wording for V4 Pro and V4.1 Flash API and chat services stands, with no cause published. Minor: DeepSeek Harness added Creator Mode (The Neuron Nov 4), generating plugins from text instructions in demos including a sticky-note board, a poster tool and a Minecraft-like game, with DeepSeek cautioning that some examples needed several rounds of refinement. No new official comment on the Oct 30 second-round financing details.
Sources: theneuron.ai
Alibaba (Qwen)
Qwen-powered Apple Intelligence developer testing in China opened Nov 4: iOS 27.3 developer beta 1 (build 24C5037i) shipped Nov 3 and testing began Nov 4 in China (Longbridge, MacObserver Nov 4), covering Live Translation in Messages, FaceTime and Phone, Writing Tools, Visual Intelligence, Clean Up in Photos, Genmoji, notification and Mail summaries, Priority Notifications and typed Siri requests, with Qwen supporting Siri, Writing Tools and Visual Intelligence for text and image understanding; users sign in with a Qwen account or Sign in with Apple, and Alibaba says it will not train on user requests or responses; eligible devices are iPhone 15 Pro / Pro Max and newer (including iPhone Air and iPhone Duo), limited to phones bought in China or devices with location and Apple Account region both set to China; the next-generation Siri AI and other WWDC 2026 features are not in this first wave, and December iOS 27.3 remains the stated general-release timing with no exact day. Qwen4 still has no official announcement or date.
Sources: longbridge.com · macobserver.com · igeeksblog.com
Moonshot AI (Kimi)
Kimi K3.1 Pro weights published Nov 4 (see Top Story 2) under the custom Kimi K3.1 Pro License with its two large-customer conditions; API pricing unchanged. No new development in the distillation accusations.
Sources: mlllm.io
Zhipu (GLM)
No new model or pricing release in the window. Single source: The Register (via The Neuron Nov 4) reports Revolut cut AI coding costs by roughly 90% after making self-hosted models its default, using self-hosted GLM models for many programming tasks with commercial models as a supplement — single-source account, not counted as a confirmed product fact.
Sources: theneuron.ai
ByteDance (Doubao)
No new proprietary model release in the window; no new model list for Doubao Work's third-party model service; no update to the 151-app SAEP refusal list.
Tencent (Hunyuan/Yuanbao)
Full Hy4 released (see Top Story 1): 770B parameters, vision and computer use, unchanged pricing, weights and system card still to come "in the coming days". The Diary of AI Nov 4 notes Hy4 is sold through routers and gateways such as OpenRouter and Vercel — a distribution observation that pairs with FlashProve's question of proving a router served the model it claims.
Sources: zhidx.com
Baidu (Ernie)
No new product fact in the window.
MiniMax
No new model release in the window. Non-product note: at the Nov 3 Wuzhen "AI Dialogue" (reported Nov 4), chief architect Feng Wen said MiniMax wants open source to give the industry something different, including engineering experience — perspectives, modalities, pitfalls, experiments and reworked tools — from its research.
Sources: 21jingji.com
Ant
Ling/Ming remain single-source reports with no official announcement text to check against.
Wuzhen AI Dialogue (other)
At the Nov 3 World Internet Conference Wuzhen Summit main forum (reported Nov 4), Zhijiang Lab director Wang Jian hosted DeepSeek research engineer Gao Xi, Alibaba ATH Qwen project lead Liu Dayiheng, MiniMax chief architect Feng Wen and International AI Governance Association chair John Higgins on why open source became a key variable in AI competition, how open models govern safety risk and earn trust, and how AI reshapes human capability; the summit theme is "Openness, Sharing and Win-Win Cooperation".
Sources: 21jingji.com
Policy and markets
US data-center ballot measures (see Top Story 3); Google's school policy (see Top Story 6). CrowdStrike announced Nov 4 the acquisitions of Fiddler AI (about $400 million) and Dreadnode (about $95 million), about $495 million combined on press-reported figures rather than company-disclosed amounts, plus an AI Kill Switch in Falcon Guardian letting administrators stop risky agents, cut network access and prevent relaunch, with endpoint-level termination still being expanded. Menlo Ventures' enterprise AI report (released Nov 3, relayed Nov 4) puts 2026 enterprise AI spending at $176.8 billion, more than five times last year, with agents generating about 80% of AI work through OpenRouter and Anthropic, OpenAI and Google taking 69% of enterprise AI spending (Menlo survey of nearly 400 business decision-makers plus public filings). Traversal raised $120 million at a reported $1 billion valuation; NICE announced about 500 job cuts (about 5% of staff) in its AI transition (both Precise Impact Nov 4).
Sources: preciseimpact.ai
Worth Reading
1."AI News Nov 4, 2026: Data Center Votes, OpenAI and Pierre" — The Diary of AI, 2026-11-04
the day's four threads — midterm data-center votes, OpenAI's Pierre acquisition, Tencent's Hy4 and Anthropic's FlashProve — useful as a cross-check for this issue's policy and acquisition lines.
Sources: thediaryofai.com
2."Everything That Happened in AI Today Wednesday, November 4" — The Neuron, 2026-11-04
the Cathedral security partnership and Reuters investigation, Google's school policy, ChatGPT's search-volume claim and Apple's 10-K agent-risk warning in one roundup, useful as a cross-check for this issue's safety and policy lines.
Sources: theneuron.ai
3."Tencent Hy4 can see and use your computer" — Zhidx (Chinese-language), 2026-11-04
hands-on testing of Hy4's native vision in programming and 3D modelling, WeChat desktop operation including the mouse-grab problem, and Ascend vLLM adaptation progress — the hands-on entry point for this issue's biggest China model release.
Sources: zhidx.com
Unverified Items
1.
Nvidia's reported 49% cap on Chinese customers: still unconfirmed in The Diary of AI's Nov 4 follow-up.
2.
Anthropic's reply to Senator Warren: due Nov 3, still no public reply visible in The Diary of AI's Nov 4 follow-up.
3.
Tencent Hy4 weights, system card and full-release licence: Tencent promises only "in the coming days" with no date, and the full-release licence is unstated; the preview's Apache 2.0 licence does not imply the full release's terms.
4.
OpenAI's reported financing round of up to $30 billion at roughly a $1.25 trillion valuation, led by UAE investors (single source: CoinMarketCap Nov 2 / XenoSpectrum relay); the Abu Dhabi agreements' official announcement discloses no amounts or financing terms.
5.
Anthropic's amended S-1 figures of $16.96 billion Q3 revenue and $895.4 billion in compute commitments (single source: AI Briefing Nov 3 summary); the amendment filing itself is confirmed, but these figures are not yet checked against the SEC text.
6.
GPT-6.2 Astra and Claude 5.5 leaks: still single source (KuCoin Nov 2, citing AIBase) with no company announcement or verifiable staged-rollout evidence.
7.
Google Gemini "hacking three companies then stopping" in an Irregular test: still only an Al Jazeera mirror (edgeone domain) relay, with no Wall Street Journal original or Google statement directly quoted.
8.
GPT-6 Astra Law pricing-page figures: still only Observatory Nov 1 and TestingCatalog relays, with no OpenAI pricing-page original checked.
9.
NVIDIA Vera Rubin bandwidth-target cuts: a September product-page revision relayed Nov 1, not a Nov 4 announcement, with no NVIDIA statement to check against.
10.
Meta–Fourier humanoid prototype: still single source (Engadget via The Diary of AI Nov 1 sidebar) with no Meta announcement to check against.
11.
Microsoft's leaked AI PC roadmap: only TechSpot Nov 4 citing a leaked slide posted by HXL on X, unconfirmed by Microsoft.
12.
Apple's engineer AI spending limit cut from $300 to $200: only The Diary of AI's Nov 4 "also noted" item citing MacRumors / Bloomberg, single source.
13.
Carried over with no new development in the window: Grok 5, Qwen4, Anthropic's IPO price range and share count (still blank), Zhipu's rumoured 3.5-trillion-parameter model, RTX PRO 5500 orders for ByteDance/Alibaba/Tencent, Ant Ling/Ming, decade-long memory contracts for Alibaba/Tencent/ByteDance, Gemini 4 Carbon's general-release date, NVIDIA Nemotron 3.5 Transcribe's official blog, distillation accusations (DeepSeek/Moonshot vs GPT-6 Astra), Meta's parental-reminder expansion, xAI's Mississippi home purchase in exchange for a dropped suit, Nvidia's reported Nscale guarantee, Grok Imagine Video 1.5 Lite, and OpenAI's actual first-round raise; Kimi K3.1 Pro weights and Apple Intelligence developer testing in China, both unverified last issue, are confirmed this issue and no longer listed here.