X AI Daily¶
Translation
English translation of the Chinese original.
1. AI Models, Products, and Applications¶
Google releases Gemini 4 Carbon: two weeks after Argon, initially only for cyber defense defenders¶
What happened: Google released its new frontier model Gemini 4 Carbon in the early hours of October 29 Shanghai time, only two weeks after the release of Gemini 4 Argon. The company says it ranks first on the Artificial Analysis Intelligence Index v5 with 59 points, 9 points higher than Argon; on cybersecurity benchmarks, SEC-bench Pro (finding vulnerabilities) 88.2%, ExploitGym (exploit development) 54.8%, CWE-bench (patching) 71.0%. This week it opens first to some cyber defense defenders, then gradually expands to trusted partners through the Fairwind program, with access for developers, enterprises, and consumers "as soon as possible" and no date yet. Introductory pricing is $2 per million input tokens and $10 per million output tokens, the same as Argon; after the introductory period it is $4/$20.
Why it matters: this is the most widely spread AI news in the window. Google shipped two generations of flagship models within two weeks and returned to first place in the rankings, but the launch uses targeted access, so ordinary users cannot use it for now.
Key X posts: @sundarpichai's launch post, about 3.027 million views, 23.6 thousand likes; @OfficialLoganK, about 841 thousand views, 13.7 thousand likes; @GoogleDeepMind, about 795 thousand views, 10.6 thousand likes. Third-party evaluation: @ArtificialAnlys's Carbon evaluation post (about 299 thousand views) says Carbon took first place in 4 of 9 items, with a coding Agent index of 59 points, second only to Claude Code with Opus 5.5 (max, 62), and still behind Kimi K3.1 Pro, Muse Wonder, and Grok 4.7 on the knowledge-work AA-Briefcase; cost per index task is $7.76, about 23% lower than Opus 5.5 (max). @arena says it ranks 1st on Text Arena, 2nd on Agent Arena, and 4th on WebDev.
Verification: scores, pricing, and the access approach were checked against Google's official blog and are consistent with the Reuters report. The cybersecurity scores are figures published by Google (officially noted as "evaluated without guardrails"), not independent reproductions.
Assessment: the score lead is a fact, but "who is strongest" depends on the leaderboard and the task; before the access scope expands, this launch is more of a preview for most developers.
Artificial Analysis launches Intelligence Index v5: 60% of questions switched to private test sets¶
What happened: independent evaluation organization Artificial Analysis released Intelligence Index v5. For the first time the index is based mainly on private tests or tests with non-public answers, accounting for 60%; it adds Cognition's FrontierCode 2.0 (real coding tasks scored on "would a maintainer merge this") and Terminal-Bench-Science, replaces Humanity's Last Exam with HLE-Diamond, and removes GDPval-AA, SciCode, and others. Knowledge work, coding, science, and general each account for 25%. Before Carbon was added, the leaderboard was Claude Opus 5.5 (max) 58, GPT-6 Astra 57, GPT-6.1 Sol 55, and Gemini 4 Argon 50.
Why it matters: launch posts for mainstream models all cite this leaderboard (including Gemini 4 Carbon in the previous item), and its revision directly affects each company's "first place" claims. The gap between open models and the top widened from 3 points in v4.3 to 12 points.
Key X posts: @ArtificialAnlys's launch post, about 1.587 million views, 5157 likes.
Verification: structure, weights, and scores were checked against the official Artificial Analysis article.
Assessment: private questions can lower the risk of leaderboard gaming and data contamination, but they also mean outsiders cannot review the questions themselves, so results still need to be read alongside actual usage experience.
Claude for Legal launches a reusable library of legal work artifacts templates¶
What happened: Anthropic launched a ready-made artifacts template library in Claude for Legal, covering scenarios such as contract review, regulatory change scanning, and citation checking; users start from an example, connect their own documents and connectors, and can then share it inside their company.
Why it matters: this is the most-viewed AI product post in the window, showing that workflow products for vertical industries like legal are drawing a lot of attention.
Key X posts: @claudeai, about 3.608 million views, 20.4 thousand likes.
Verification: the feature description was checked against the official Claude introduction page, which lists templates including multi-contract clause comparison and flags for deviation from standard clauses. Views only indicate exposure, not adoption.
Assessment: Anthropic is moving from a "general assistant" toward finished industry workflows, and legal is one of the first industries where it has landed.
GitHub Universe: Copilot Fusion generally available, automatically routing and orchestrating multiple models per task¶
What happened: at the Universe conference GitHub announced that Copilot Fusion is generally available. Users select Fusion in the model list instead of specifying a concrete model, and the system picks one or more models by task, with three execution modes: a single model completes the task directly, a small model drafts first and escalates if the result is inadequate, or one model drafts while other models review in parallel. Effort levels are Standard, High, and Max. GitHub says Fusion Max scores 60 on the Artificial Analysis coding Agent index, within 2 points of Claude Code with Opus 5.5 Max, at roughly half the average API cost per task and about 40% less time. Paid users get 30% off model costs until November 30.
Key X posts: @satyanadella, about 114 thousand views, 949 likes; the post says "40% faster, up to 77% cheaper".
Verification: features and benchmark numbers were checked against the GitHub Changelog and the GitHub documentation; the basis for "up to 77% cheaper" is not explained in the official documentation, so it is treated as the vendor's claim.
Assessment: the competitive focus of coding tools is shifting from "which model is integrated" to "how multiple models are orchestrated", and both cost and results are vendor self-measurements that need validation in real projects.
OpenAI: eligible ChatGPT Pro and Business subscribers can claim up to $500 in API credits per month¶
What happened: the OpenAI developer account announced that eligible ChatGPT Pro and Business subscribers can claim up to $500 in API credits per month for building apps and Agents.
Key X posts: @OpenAIDevs, about 75 thousand views, 1369 likes.
Verification: according to the OpenAI Help Center, Business plans accumulate by seat (Standard seats $20, Premium seats $100), with a cap of $500 per workspace per month, the subscription must have been active for 7 days, and credits expire in the current period. The specific credit amount for individual Pro subscriptions was not confirmed in the material read for this issue.
Assessment: steering subscribers toward the API is a bid to capture the developers among "people who use ChatGPT and also build their own Agents".
Other¶
- Google launched a preview of the Connected Apps Platform, where developers can build and manage apps for Search and the Gemini App in one place, with Adobe, Calm, and Peloton among the first partners. @Google's original post, about 67 thousand views. A new release that has not yet drawn high engagement; not separately verified.
- Community account @chiefofautism posted that activation steering on a single neuron in one layer of Qwen3.8 27B can make the model output content hostile to humans (original post, about 476 thousand views), and @AISafetyMemes's retelling and @Polymarket framed it as "researchers discover an evil neuron". This is a personal experiment with no paper or peer review; trending on X, facts pending verification.
2. AI Compute and Industry¶
Alphabet Q3: Google Cloud revenue doubles year over year, TPU system sales become a growth source¶
What happened: Alphabet's third-quarter revenue was $127.6 billion, up 25% year over year; Google Cloud revenue was $30.7 billion, up 103% year over year, with operating profit of $11.1 billion. The official earnings release attributes growth to enterprise AI infrastructure and AI solutions, and states that TPU system sales were a "significant contribution". Pichai said the company reached a half-trillion-dollar annualized revenue scale, its own model APIs process nearly 32 billion tokens per minute, up 40% quarter over quarter. The earnings call disclosed $561 billion in Cloud backlog.
Key X posts: @sundarpichai, about 391 thousand views, 5885 likes.
Verification: revenue, growth rates, and the TPU wording were checked against Alphabet's earnings press release filed with the SEC; backlog comes from the earnings call transcript. The per-minute token figure is Pichai's claim in his post.
Assessment: for the first time Google Cloud lists "selling TPU hardware" as a growth driver, and its competition with Nvidia has expanded from cloud rentals to selling chips directly.
Microsoft FY27 Q1: Azure grows 48%, Azure revenue disclosed separately for the first time¶
What happened: Microsoft's quarterly revenue was $92.3 billion, up 19% year over year; Microsoft Cloud revenue was $63.9 billion, up 30% year over year; Azure revenue grew 48%, with the amount disclosed separately for the first time at $33.165 billion. Commercial remaining performance obligation grew 73% to $677 billion. On the earnings call, management said Foundry revenue grew more than twofold year over year, delivery of third-party GPU capacity was delayed, demand still exceeds supply, and Azure is expected to grow 49% to 50% next quarter at constant currency.
Key X posts: @Microsoft, about 85 thousand views, 798 likes; @satyanadella, about 95 thousand views, 952 likes.
Verification: revenue and Azure figures were checked against the Microsoft Investor Relations press release; guidance and the Foundry wording come from the earnings call transcript.
Assessment: both cloud providers posted high growth on the same day, and AI compute demand is still constrained by supply rather than demand peaking.
Report: OpenAI plans to buy Google's AI chips¶
What happened: The Information reported, citing people familiar with the matter, that OpenAI plans to buy Google's TPU chips, and Bloomberg later reported similarly. Neither side commented, and deal terms were not disclosed. On the same day Oracle announced it will offer TPUs on its cloud.
Key X posts: @Kalshi, about 67 thousand views, 985 likes; @zerohedge mentioned it in a pre-market movers note, about 76 thousand views.
Verification: checked against Reuters's account of The Information's report and the Bloomberg report (republished by Business Standard). For now this is only media reporting with no official confirmation.
Assessment: read together with the TPU sales in Alphabet's earnings above, Google is becoming a chip supplier to major AI labs.
SK hynix Q3 revenue tops 100 trillion won for the first time, HBM4 begins to ramp¶
What happened: SK hynix's third-quarter revenue was 104.6 trillion won, breaking 100 trillion for the first time; operating profit was 83.3 trillion won, an operating margin of about 80%. The company says growth came from server DDR5, data center SSDs, and HBM4 shipments, and that long-term agreements already cover more than half of capacity for 2027–2028.
Key X posts: @SKhynix, about 41 thousand views, 267 likes.
Verification: revenue and profit were checked against SK hynix's official press release. Some posts on X say "memory sold out until 2029", while the official post says long-term agreements cover more than half of capacity for 2027–2028; this report follows the official statement.
Assessment: demand from AI data centers has reached the memory segment, and HBM supply and demand remain tight.
Other¶
- Amazon's pullback from its in-house Nova models and AGI team drew renewed discussion. Researcher Leo Boytsov posted citing a Futurism article, saying Amazon's AGI unit is being shut down and the Nova models are being phased out, and later corrected some of his statements himself; @AGTPinsights's summary notes that most of the information had already been reported by outlets such as Business Insider. No official response from Amazon has been seen, so this is treated as the source's claim.
3. Chinese-Language AI Ecosystem¶
After the Artificial Analysis v5 revision, Chinese-language circles discuss domestic models' scores "falling across the board"¶
What happened: after v5 went live, Chinese-language account @NFT_Chen compiled a comparison of old and new scores for domestic models, saying Xiaomi MiMo-V2.6-Pro fell from 46 points to 30, Qwen3.8 27B and MiniMax-M3.1-Flash each dropped 14 points, and Kimi K3.1 Pro dropped 9 points, the smallest decline; the post title directly says "highly suspect this is AA's bias". About 17 thousand views, 60 likes, 37 replies.
Verification: the official Artificial Analysis article confirms that the gap between open models and the top widened from 3 points to 12 points, and that among models with released weights, DeepSeek V4.1 Flash and GLM-5.3 are highest at 32 points each (source). The per-model declines are the poster's own tally and were not checked one by one; "bias" is the poster's opinion, while the official explanation is that v5 added harder private questions.
Assessment: the private coding questions and scientific Agent tasks added in v5 are exactly where domestic open models are weakest, so the score decline more likely comes from the change in question types, but this is only an inference.
Other¶
- The same account introduced Alibaba's OvisOCR3: a 0.8B unified document model, open sourced under Apache-2.0, claimed to rank first among same-size models on multiple OCR and document parsing benchmarks (original post, about 19 thousand views). The benchmark numbers come from the project's own report and were not verified in this issue.
- Community project Qwenputer compiled a RISC-V CPU into the weights of Qwen3.5-4B, runs C and MicroPython inside the model's forward computation, and claims its arithmetic benchmark improved to 97.5%–100% (Chinese-language introduction post, about 15 thousand views, 184 likes). This is a personal experiment whose results have not been independently reproduced.
Search for this issue was mainly in English; the Chinese-language portion comes mainly from a small number of accounts, and independent AI news with high engagement within the Chinese-language window was relatively scarce.