AI Daily Brief · Oct 29, 2026 | Google launches Gemini 4 Carbon; Alphabet and Microsoft earnings: cloud surges, Alphabet falls ~10% after hours; OpenAI's Dot takes texts; DOJ looks at Nvidia's Hugging Face deal

Google owned both the model news and the earnings news over the past 24 hours. Gemini 4 Carbon edged past Claude Opus 5.5 on Artificial Analysis Intelligence Index v5 with a score of 59 and leads on cybersecurity evals, but it is rolling out first to cyber defenders only. The same day, Alphabet’s Q3 cloud revenue doubled while search revenue missed estimates for a second straight quarter, sending shares down about 10% after hours; Microsoft’s Azure grew 48%. OpenAI let Pro users text their Dot and added API credits to Pro and Business subscriptions. The Information reported that DeepSeek and Moonshot heavily distilled GPT-6 Astra, while Nvidia faced DOJ questions over its Hugging Face deal and a New York Times spy report. Anthropic launched a library of legal artifacts for Claude, and Google’s AI Mode drew fire for dodging the question of who won the 2020 election.

Today’s highlights

• Google released Gemini 4 Carbon: No. 1 on AA Intelligence Index v5 at 59, 88.2% on SEC-bench Pro, first available to trusted cyber defenders, with introductory pricing of $2/$10 per million tokens.
• Alphabet’s Q3 cloud revenue hit $30.7B, up 103%, but an overall miss sent shares down about 10% after hours; Microsoft’s Azure grew 48% and was disclosed in dollars for the first time at $33.17B.
• OpenAI: US Pro users can now text their Dot over iMessage/RCS, and Pro plans come with $100–$500 a month in API credits.
• The Information reported that DeepSeek and Moonshot heavily distilled GPT-6 Astra to train new models [Media report].
• The DOJ is asking about Nvidia’s $12.93B purchase of Hugging Face; Nvidia also released Nemotron 3.5 Transcribe, an open speech-recognition model.

1. Google launches Gemini 4 Carbon: No. 1 on AA Intelligence Index v5, leads on cyber evals, staged rollout to defenders first

Launch: Sundar Pichai called Gemini 4 Carbon Google’s “next frontier model,” arriving just two weeks after Argon. It ranks first on Artificial Analysis Intelligence Index v5 at 59 and scores 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (exploitation) and 71.0% on CWE-bench (patching) (31.5K likes, ~5.51M views). Logan Kilpatrick and Google DeepMind posted the launch at the same time (15.3K / 13.3K likes). Elon Musk replied “Impressive” (28.2K likes, ~3.61M views).

Safety and alignment: According to Google’s blog, Carbon’s successful-cheating rate on Terminal-Bench 4.0 is 0.0% with 0.4% attempted cheating; high-severity misalignment alerts fell by almost 80%; unsupported claims were roughly halved; overreach fell 26% and under-service 29%; and the Gray Swan indirect prompt-injection attack success rate was 0.0% (Argon: 0.7%) [Company claim].

Access and pricing: A subset of trusted cyber defenders gets access this week, followed by more partners through the Fairwind Program and then wider availability. Google says this is part of the US voluntary pre-release testing process. Introductory pricing is $2/$10 per million input/output tokens with 95% off cached input, rising later to $4/$20.

Third-party evals: Artificial Analysis put Carbon first on v5 at 59 (the previous leader, Opus 5.5, scored 58). It scored 59 on the Coding Agent Index, behind Claude Code + Opus 5.5 (max) at 62. Cost per Intelligence Index task was $7.76 versus $10.04 for Opus 5.5 (our calculation: about 23% lower) and $6.25 for GPT-6 Astra. It scored 54 on AA-Omniscience (Opus: 48), but its hallucination rate rose from 4% to 11%; on AA-Briefcase it scored 1609, behind Grok 4.7, Muse Wonder and Kimi K3.1 Pro (4.7K likes) [Self-reported by the leaderboard]. @testingcatalog said Carbon (High) took first on Text Arena with 1528 points, second on Agent Arena (behind Opus 5.5) and fourth on WebDev Arena (389 likes); Google’s blog says it gained 89 Elo on WebDev Arena to 1768.

Reactions: @Austen said every new Google model crushes the benchmarks, yet he knows almost no one who uses Gemini regularly (2.3K likes) [Opinion].

Sources: @sundarpichai 1 · @OfficialLoganK · @GoogleDeepMind · @ArtificialAnlys 1 · @testingcatalog · @elonmusk 1 · @Austen · Google blog

2. Alphabet and Microsoft report on the same day: Google Cloud doubles but Alphabet misses and falls ~10% after hours; Azure grows 48%

Alphabet Q3: Revenue was $127.6B, up 25%. Google Cloud brought in $30.709B, up 103% from $15.157B a year earlier; Search & other was $63.35B, up 12%; operating margin was 31.8% and EPS $2.12. On the call, management cited capex of $51.4B, a Cloud backlog of $561B (just over half expected to be recognized within 24 months), 32B tokens processed per minute (up 40% quarter over quarter and more than 4x year over year), and 1B Gemini app users. Pichai posted the results (10.6K likes, ~806K views).

Market reaction: CNBC called it an across-the-board miss: search revenue came in below estimates for the second straight quarter, free cash flow was negative $9B, and the stock fell about 10% in extended trading. Alphabet kept its full-year capex guidance of $195B–$205B and said spending will rise significantly in 2027. Analysts at UBS, Mizuho and D.A. Davidson said the results put fears of AI cannibalizing search back in focus [Opinion].

The “RSI” remark: Pichai said on the call that he expects continued acceleration over the next 12 months because of RSI (per the Motley Fool transcript). @kimmonismus read it as Pichai officially declaring RSI (recursive self-improvement) (2.1K likes) [Opinion].

Microsoft FY27 Q1: Revenue was $92.3B, up 19%. Microsoft Cloud reached $63.9B, up 30% (31% in constant currency); Azure and other cloud services grew 48%; commercial remaining performance obligation was $677B, up 73%; net income was $36.6B. CNBC said Microsoft disclosed Azure revenue in dollars for the first time: $33.17B. Satya Nadella posted the results (8K likes, ~1.23M views).

Sources: @sundarpichai 2 · @kimmonismus 1 · @satyanadella 1 · Alphabet earnings release (SEC Exhibit 99.1) · Pichai’s earnings remarks · Motley Fool call transcript · CNBC (Alphabet earnings) · CNBC (search and market reaction) · Microsoft press release · Microsoft Investor Relations · CNBC (Microsoft earnings)

3. OpenAI: Text your Dot; API credits for Pro and Business plans; Chat/Work merger “not rolling out just yet”

Texting your Dot: The ChatGPT account announced that users can now text their Dot (6.7K likes, ~1.77M views). Per OpenAI’s docs, it is a limited beta for US ChatGPT Pro users over iMessage on iPhone and RCS on Android; users can send text, photos and voice notes, and the Dot can proactively text updates. Setup requires choosing Texting on the Dot profile and verifying a phone number. In a “Day 24” post, @thsottiaux showed forwarding, voice notes and texting a Dot through Siri (3.4K likes, ~911K views).

API credits with subscriptions: OpenAI Developers announced API credits for eligible ChatGPT subscriptions (7.7K likes, ~1.56M views): Pro plans get $100, $200 or $500 a month by tier; Business plans get $20 per paid Standard seat or $100 per paid Premium seat, pooled up to $500 per workspace. The help page adds that the subscription must have been active for 7 days and that credits do not roll over.

Chat/Work clarification: @thsottiaux said the team is working on bringing Chat and Work together but it’s “not rolling out just yet”; the plan is to do it at some point this year, carefully (5.1K likes, ~1.16M views).

Also: @kimmonismus, citing another post, said Dots would be available in the EU that day for business and enterprise, with consumer plans soon (396 likes) [One-sided claim]. OpenAI expanded its Oracle partnership: US Oracle customers can buy the API, Codex and ChatGPT Work through Oracle Marketplace with eligible Oracle Universal Credits (official). @derrickcchoi’s post also said “130K ChatGPT + 95K+ Codex users now,” without saying what the figures cover (163 likes) [One-sided claim].

Sources: @ChatGPT · @OpenAIDevs 1 · @OpenAIDevs 2 (448 likes) · @thsottiaux 1 · @thsottiaux 2 · @kimmonismus 2 · @derrickcchoi · ChatGPT Learn: Text your dot · OpenAI Help: Getting started with your dot · The Verge · OpenAI Help: API credits · OpenAI developer docs · OpenAI: Oracle partnership

4. The Information: DeepSeek and Moonshot heavily distilled GPT-6 Astra to train new models [Media report]

The report: The Information reported on Oct 28 at 6 p.m. (US Pacific time) that GPT-6 Astra has become Chinese AI labs’ top distillation target, with DeepSeek and Moonshot relying heavily on its outputs. Moonshot’s Kimi K3.1 Pro, released last week, and DeepSeek’s V4.1-Flash 1017, released Oct 17, both benefited. Reuters said neither company responded to requests for comment [Media report, not confirmed by the company].

Background: Reuters noted that OpenAI’s Sept 30 post linked a July distillation campaign to individuals associated with Moonshot, and that Anthropic says its countermeasures cut distillable data by about 90%–99%.

Reactions: @kimmonismus said the damage to Kimi’s reputation keeps growing (1.2K likes) [Opinion]; @MTSlive relayed the report (510 likes); @bindureddy argued Chinese labs moved to distilling Astra because Anthropic blocked them (337 likes) [Opinion].

Sources: @kimmonismus 3 · @bindureddy · @MTSlive · The Information · Reuters (via Investing.com) · TNW

5. Nvidia: DOJ asks about the Hugging Face deal; NYT says a Jensen Huang contact is believed to be a Chinese intelligence asset; open speech model Nemotron 3.5 Transcribe

DOJ inquiry: [Media exclusive] Reuters reported exclusively on Oct 28 (US Eastern time) that the US Justice Department is asking about Nvidia’s $12.93B purchase of Hugging Face, including whether Nvidia would gain control of a fundamental distribution channel for AI. The inquiry is part of a broader DOJ look at Nvidia’s acquisitions, investments and sales practices that has also covered SchedMD, Groq and Run:ai. The deal is expected to close in the first half of 2027. The DOJ and Nvidia declined to comment. @zerohedge and @MarioNawfal relayed the news (263 / 216 likes).

Spy report: [Media report] The New York Times reported on Oct 29 that US officials believe Alice Huang (Huang Le), a Megaspeed executive close to Jensen Huang, is a Chinese intelligence asset, and that President Trump was briefed about a year ago but took no action. Nvidia said the meetings were ordinary customer interactions and no sensitive information was shared; the Chinese embassy dismissed the report as groundless smears. The Verge and Tom’s Hardware followed up. @IntCyberDigest (2.9K likes, ~343K views) and @Kalshi (1.5K likes) relayed it.

Nemotron 3.5 Transcribe: Nvidia released an open speech-recognition model; its post describes a 30B MoE with about 4B active parameters supporting 40 languages (7.7K likes, ~754K views). The Hugging Face model card lists 34B total parameters (4.3B active) including the audio encoder, a Mamba-2–Transformer hybrid MoE, recordings up to 1 hour, and the openmdw-1.1 license. Artificial Analysis said it scores 2.8% AA-WER, tied for first among open-weights models with Voxtral Small (563 likes) [Self-reported by the leaderboard].

Sources: @NVIDIAAI · @ArtificialAnlys 2 · @zerohedge · @MarioNawfal · @IntCyberDigest · @Kalshi · Reuters (via Investing.com) · Reuters (via U.S. News) · The New York Times · The Verge · Tom’s Hardware · Hugging Face model card · Hugging Face blog

6. Anthropic launches a legal artifacts gallery for Claude for Legal: ready-made cite checks, regulatory scans and review templates

Launch: The Claude account announced a gallery of artifacts for legal work in Claude for Legal (28.4K likes, ~6.33M views). Anthropic’s article lists ready-made artifacts including Cite check, a jurisdiction survey matrix, a matter board, a playbook deviation board, a regulatory horizon scan and tabular review. It connects to 60+ legal connectors (such as Bloomberg Law, Filevine, Ivo, Legora and Patlytics), with partner artifacts from Goodwin, Harvey, Foley & Lardner and PwC, plus nine access-to-justice artifacts (five built by nonprofits) and a new Legal Ambassadors Program. Claude for Legal launched in May.

Reactions: @BoringBiz_ posted a meme about lawyers watching Claude automate half their job in a single release (8.8K likes, ~684K views); @ai_for_success asked how many small legal-AI startups had just been put out of business (115 likes) [Opinion].

Sources: @claudeai 1 · @claudeai 2 (788 likes) · @ai_for_success · @BoringBiz_ · Claude article

7. Google AI Mode won’t say who won the 2020 election; Google blames a “technical gap” it is fixing

What happened: Author Charles Mann said he reproduced the behavior himself: two queries a minute apart, identical except for one digit of the year, and AI Mode dodged only the 2020 one (7.6K likes, ~924K views). @CameronCorduroy called it “definitely some weird shit going on” (15.8K likes, ~1.45M views).

Media checks: Ars Technica reported that AI Mode declined to name the 2020 winner and instead pointed users to Federal Election Commission results. The behavior varied by user, asking again sometimes produced a direct answer, and tests through foreign VPNs got direct answers.

Google’s response: As relayed by the press, a Google spokesperson said the company is fixing the issue with 2020 election queries in AI Mode, describing it as a technical gap affecting 2020, one that also affects 2022 Senate races; The Verge reported that Google is looking into it.

Reactions: Climate scientist Zeke Hausfather argued Google had deliberately stopped its AI from answering, calling that kind of political intervention in LLM responses deeply problematic (2.1K likes) [Opinion], which differs from Google’s technical-gap explanation.

Sources: @CharlesCMann · @CameronCorduroy · @hausfath · Ars Technica · The Verge

In brief

• GitHub Universe: Copilot Fusion: GitHub launched Copilot Fusion to route and orchestrate multiple models, saying Fusion Max came within two points of Claude Code + Opus 5.5 Max on the Artificial Analysis Coding Agent Index at about half the API cost per task (@github, 544 likes); Satya Nadella said Fusion is “40% faster and at up to 77% lower cost,” and highlighted the unified Copilot harness coming to Foundry and Copilot Projects for multi-agent orchestration (@satyanadella 2, 1.4K likes) [Company claim]. The “up to 77%” figure comes from GitHub’s own evaluation. Official: GitHub changelog · GitHub blog
• Google Connected Apps Platform preview: Developers get one place to build and manage apps for Google Search and Gemini, built on the open MCP standard, with launch partners including Adobe, Calm and Peloton (@Google, 796 likes); on Android, Google also launched Android MCP powered by AppFunctions (@MishaalRahman, 1.1K likes). Official: Google blog · Android Developers Blog
• Cartesia’s new H-Net (byte-level language models): Albert Gu said the new hierarchical networks deliver up to 3–5x efficiency gains in controlled scaling experiments and, at larger scale, match an open model trained on 11T tokens with about a tenth of the data and compute (@_albertgu, 6K likes, ~685K views) [Company claim]. Official: Cartesia blog
• New NanoGPT speedrun record: Herman Brunborg said retrieval-augmented n-gram hints cut the record from 39.9s to 6.7s (−83.2%), more than all records since 2025 combined (@hermanbrunborg, 2.4K likes) [Self-reported, not independently verified]. Write-up: author’s blog
• Particle-1: Suhail Doshi’s new company Particle released its first developer-focused model, claiming parity in performance while using about 45% fewer tokens (@Suhail, 2K likes) [Company claim]. Official: Particle blog
• Rime Prism speech-to-speech model: Lily Clifford said Prism has a 300ms response time, 4x faster than OpenAI GPT-Live (@lilyjclifford, 965 likes, ~727K views) [Company claim]; Rime’s own product page, however, lists time to first audio of about 0.57 seconds, so the two figures don’t match. Official: Rime
• Deepgram acquires Daily: Daily, maker of the open-source Pipecat voice-agent framework, is joining Deepgram; Pipecat stays open source under an independent Pipecat Foundation (@DeepgramAI, 21 likes). Coverage/official: TechCrunch · Daily blog
• Short seller targets Cerebras [One-sided claim]: Culper Research said it is short Cerebras ($CBRS), arguing its wafers break down after about two years of use, a lifespan that undermines the business (@CulperResearch, 1.3K likes, ~482K views). Report: Culper report (PDF)
• Qwen4 “next week”? [Unverified]: Posting from Amsterdam, @jtdavies said Alibaba Cloud’s CTO said Qwen4 is coming “next week” (@jtdavies, 917 likes); Alibaba has not announced a release date.
• Kimi connects to Google apps: Kimi now works with Gmail, Google Drive and Google Calendar via OAuth (@Kimi_Moonshot, 357 likes).
• A “points system” for Claude bans? [Unverified]: @make_claude posted a long thread claiming the Chinese community had reverse-engineered a points-based Claude ban system, listing factors such as signup, payment, IP and devices (@make_claude, 11.3K likes, ~1.42M views); Anthropic has not commented.
• Early Neuralink speech-restoration data: Neuralink said one participant with ALS reached 75 words per minute at a 5.8% word error rate on a 125K-word vocabulary in their first week (@neuralink, 8.6K likes, ~936K views) [Company claim].
• Damodaran values Anthropic: Valuation expert Aswath Damodaran said Anthropic filed for its IPO on Oct 26 and that he values its equity at about $1.475 trillion (@AswathDamodaran, 4.1K likes, ~920K views) [Opinion].
• Lawsuit over ChatGPT medication advice [Allegation by one party to the lawsuit]: The New York Post reported that a 31-year-old woman died after tapering off her epilepsy medication on ChatGPT’s advice; her family has sued OpenAI and Sam Altman (@nypost, 9.6K likes, ~1.14M views). Coverage: New York Post · TechCrunch
• Grok Bot 0.78.1: @ericzakariasson said the 0.78.1 stable release is rolling out, with shopping results as product cards, new subscription templates and Shop Pay in the vault (@ericzakariasson, 1.1K likes) [Company claim]; @jamesdouma said Amazon locked out his Grok Bot, so it bought the same item from a different site at a lower price (@jamesdouma, 1.8K likes, ~1.62M views) [First-person account]; Elon Musk replied, “That’s what I do too” (@elonmusk 2, 15.5K likes).
• Musk: saving is pointless: Elon Musk replied that saving money for the distant future is pointless with the advent of superintelligent abundance (@elonmusk 3, 12K likes, ~1.16M views) [Opinion].

---

About this column and how we verify

1. AI-assisted editing: AI Daily Brief is AI-assisted; this issue was produced with help from Grok Bot and covers Beijing time Oct 28, 2026, 22:15 through Oct 29, 2026, 22:09.
2. Tweet-level checks: Every cited post was fetched to confirm it exists, the author matches, and the timestamp falls in this period; all selected posts were also re-read via the X API. Like and view counts are as of editing time (around 22:11 Beijing time on Oct 29).
3. X News summaries are leads only: X News blurbs are Grok-generated and used only to find stories; facts are checked against original posts, official pages and mainstream coverage. When a summary conflicts with a primary source, the primary source wins. Figures marked “our calculation” were computed by us from public data.
4. Items about Grok Bot: The “Grok Bot 0.78.1” item in brief concerns the product used to help produce this column; it was included only on the strength of its engagement, under the same standards as every other item.
5. Label key: [Unverified] = circulating without an official or other primary confirmation; [Media report] = from press coverage that may differ in framing from the official release; [Media report, not confirmed by the company] = press report the companies involved have not confirmed; [Media exclusive] = a single outlet’s exclusive, citing people familiar; [Company claim] = from a company release or official account, not independently reproduced; [Self-reported by the leaderboard] = from a benchmark publisher’s own write-up; [One-sided claim] = from one party only, with no other source corroborating; [Self-reported, not independently verified] = results published by the person who achieved them, not independently reproduced; [Allegation by one party to the lawsuit] = a claim in a complaint, not yet decided by a court; [First-person account] = someone describing their own experience; [Opinion] = judgment, not a factual claim.