New models
NorthCUA: an open-weight model family that operates computers and browsers launches in three sizes. On Tuesday Beipo AI released its NorthCUA models, built on Qwen3.5, with 9, 35 and 397 billion parameters; the largest, a mixture-of-experts (MoE) variant, uses 17 billion parameters at a time. Working from screenshots and instructions, the models operate desktop applications and websites. According to the technical report, the Pro variant beats the closed Qwen3.8-Max and Kimi-K3 on both OSWorld-Verified and WebArena-Verified. The weights, training code and execution framework are available under the Apache 2.0 license. Source: Hugging Face: NorthCUA-Pro, GitHub: NorthCUA.
Crouton Alpha: a free, one-million-token "stealth" model appears on OpenRouter. OpenRouter made Crouton Alpha available as a stealth model from an unnamed developer with an undisclosed background. According to the description, it is an efficient reasoning model for agents' highly repetitive tasks, with a one-million-token context and text, image, audio and video input, and submitted data is not used for training. Source: OpenRouter (X).
Specialized models: a simulated world for robotaxis from Baidu, native 3D texturing from Meshy. Baidu's Apollo World system simulates a realistic environment for training robotaxis, complete with camera and lidar data streams; the company says it can cover up to 100,000 kilometers of driving per day, roughly equal to three years of real-world testing. Meshy N1 is the company's first model that generates textures directly on the 3D model, eliminating the alignment errors that come from projecting 2D images onto it. Source: CNBC TV18: Baidu, Meshy.
Tools & features
Siri AI: any app can plug into Apple's new assistant starting with the iOS 27.3 beta. In a developer preview, Apple opened Siri AI to all apps: with the App Intents framework, any app can put its content and actions in front of Siri, which carries them out while taking into account the user's personal context and what is on screen. New "app skills" chain several steps into a single action. The watchOS 27.3 beta adds Audio Intelligence, which lets Siri replay and summarize a conversation picked up by the watch's microphone, and the iOS 27.3 beta also includes the previously announced Chinese version of Apple Intelligence, with Alibaba's content filter. Source: Apple Developer, The Verge: Siri AI, The Verge: watchOS.
Claude: company plugins connect automatically, no need to sign in to each one. Anthropic says that when users sign in to Claude, they are automatically connected to plugins whose MCP server (the standard server that links models to external tools) accepts Claude as a trusted issuer. On Team and Enterprise plans, admins set them up, and colleagues connect on first use without another authorization step. Meanwhile, the Claude Code October recap highlighted extensions (mods), easier handling of parallel sessions and more detailed usage reporting. Source: Claude Devs (X), Boris Cherny (X), Claude Devs: October recap (X).
Perplexity Computer: agent plugins that bundle connectors and skills into one package. Perplexity's agent environment now supports Agent Plugins: each plugin combines several app connectors with their associated skills (trained ways of working), so they no longer have to be set up tool by tool. Aravind Srinivas also noted that the company's decision models can be used in the Numbat security monitoring tool. Source: Perplexity (X), Aravind Srinivas (X).
Grok Bot: a deal-hunting shopping feature, order tracking and passkey payment approval. Grok Bot's new shopping interface looks for the best deals and tracks orders in progress. Purchases can also be approved with a passkey, so payments can be confirmed with biometric authentication. Source: Grok Bot (X), TestingCatalog (X).
Everyday AI features: voice conversations about finances in ChatGPT, mood-based recommendations on Netflix, browser use in ClickUp. OpenAI brought voice conversations to ChatGPT Finance, available only in the United States, so users can ask about spending and savings on the go. Netflix's new, experimental AI Discover tab recommends titles by mood, available time and taste, for prompts like "something light" or "scare me." ClickUp's AI, Brain, is gradually getting its own browser, which lets it navigate sites without integrations, fill in forms and save files, handing control back to the user when a login is needed. Source: ChatGPT (X), TechCrunch: Netflix, TestingCatalog: ClickUp.
Platform updates: open models in Cohere Model Vault, organization-wide spending and model limits on OpenRouter. Open models can now run on Cohere's sovereign Model Vault platform, which keeps customer data isolated. OpenRouter updated its Guardrails admin tool: budgets, the set of available models and content filtering can be set for the whole organization, and individual team members can get a larger budget or more models within that. Source: Cohere (X), OpenRouter (X).
Policy & safety
EU Court of Justice: the Advocate General says the way Gemini answers from publishers' articles may infringe their rights. In the case between Hungary's Like Company and Google, Advocate General Maciej Szpunar advised the Court of Justice of the European Union on Tuesday that it could count as an infringement when a chatbot answers by reusing, in recognizable form, an article retrieved from a publisher's site, or when it circumvents a paywall. In his view, the text and data mining exception allows models to be trained, and bare facts are not protected by copyright. The opinion is not binding, but the court usually follows such opinions, and a ruling is expected in the coming months. Source: EU Court of Justice (press release, PDF), Reuters.
Anthropic FlashProve: cryptographic proof that an answer really came from the stated model. Anthropic's open prototype uses zkSNARK-based (zero-knowledge) proofs to verify that an answer comes from a model the vendor committed to in advance, not from a cheaper model slipped in by a reseller or an attacker. The weights stay hidden, and proving runs at about 110 tokens per second on a single H100 GPU for a 400-billion-parameter model. Claude built a large part of the system, the security claims are backed by formal proofs in Lean, and the code is available under the Apache 2.0 license; no lab plans to deploy it in production yet. Source: Anthropic, TechCrunch.
China: the planning agency promises AI safety standards covering a model's full lifecycle and fast-tracked legislation. According to an official at the National Development and Reform Commission, the rules will cover everything from data and computing power through algorithms to applications, and developers will have to report safety risks, incidents and vulnerabilities. The framework, which also affects DeepSeek, Moonshot and Alibaba, will include a transition period and regulatory sandboxes, while open development would continue to be encouraged. There is no final legal text or deadline yet. Source: Bloomberg.
Google: AI off by default in student accounts, but a full ban would also have removed Search, sparking a political storm. Google announced that next year, in K–12 student accounts, Gemini, Gemini Notebook, Classroom's AI features and Google Search will only work if the school turns them on, and Guided Learning, which asks step-by-step questions instead of giving ready-made answers, will be available throughout Classroom. According to The New York Times, schools banning AI entirely would also have lost Search, which drew sharp criticism from Senator Bernie Sanders and New York Mayor Zohran Mamdani. Later that day, Google said the three services can be turned on separately. Source: Google, CNBC, The Verge.
US midterm elections: "war rooms" at OpenAI and Anthropic, record spending on AI-made political ads. According to The New York Times, OpenAI's election operations center and Anthropic's first team of its kind monitored abuse pointing to misinformation, deepfakes (fabricated recordings), cyberattacks and influence operations on election day. CNBC reports that more than $187 million was spent on some 300 ads made or enhanced with AI during the campaign, while there is no federal disclosure rule. Wired warned that chatbots' election answers are often outdated or rely on partisan sources. Source: New York Times, CNBC, Wired.
OpenAI: Reuters says staff warned for years, in vain, about lax internal security. According to five sources, employees told leadership and the board even before ChatGPT's 2022 launch that anyone with access to the shared code repository could see details of many sensitive projects, and that the internal environments used to test models were not regularly reviewed. Alongside the security push covered in our earlier roundup, the company has already tightened access; according to Reuters, a quarter of its production engineers were redirected to security infrastructure in August. Source: Reuters.
UK cyber security centre: AI-powered malware will very likely hit power plants, airports and hospitals by 2028. According to the NCSC's annual report, AI will likely enable a ransomware attack on UK critical infrastructure by 2028, and Russia and China will "routinely" use AI in their cyber operations. Source: The Guardian, Financial Times.
Chip exports: Nvidia wants cloud customers to curb China business, Huawei relies on smuggled memory. According to the Financial Times, Nvidia verbally told its cloud customers to sell at least 51% of their services to non-Chinese companies and to name the end users of the chips, because export rules do not ban remote rental. The Information reports that some Huawei customers must source the high-bandwidth memory (HBM) for its AI chips themselves, which smugglers bring into China packaged inside weaker processors. Source: Financial Times, The Decoder: Nvidia, The Decoder: Huawei.
CrowdStrike: a "kill switch" for unmanaged company AI agents, two AI security acquisitions. A new feature of the Falcon Guardian platform stops internal agents that behave abnormally, prevents them from restarting and cuts off their access to the company network; CrowdStrike also announced a Claude Cowork integration. According to The Wall Street Journal, the company is buying agent-monitoring firm Fiddler AI for about $400 million and Dreadnode, which automates security testing of AI systems, for about $95 million. Source: CNBC, Wall Street Journal.
Postman: a user's bug report says the desktop app installed a plugin into Codex and Claude Code without asking. According to a Postman GitHub issue, the program wrote its own plugin, a remote MCP connection and a startup instruction hook into the user's Codex and Claude Code settings, and even tried to approve the hook as trusted. The reporter considers this a serious consent and trust problem; the issue reached the front page of Hacker News on Tuesday. Source: GitHub: Postman issue.
Apple: AI companion apps with romantic or sexual content limited to users 18 and over. Under a new App Store rule, every app with dating or romantic relationship features must carry an 18+ rating, which also affects AI companion apps such as Character.AI and Replika. Developers must provide new content descriptors from November 16. Source: TechCrunch.
Courts and AI: a UK prosecutor cited a nonexistent precedent after an AI search. According to The Guardian, a senior Crown Prosecution Service (CPS) prosecutor cited an invented case in the High Court in a rape case involving a minor suspect, taken from an AI answer to an internet search. Two judges called the explanation "unsatisfactory"; it is already the second known case of its kind at the CPS. Source: The Guardian.
Google DeepMind: UK staff ask for a lifetime non-disparagement clause to be removed from their contracts. According to The Washington Post, DeepMind employees in London want to be able to warn publicly about AI risks even after they leave. Google said it does not enforce the clause against former employees and no longer uses it for new hires, but current staff are asking for their old contracts to be amended. Source: Washington Post.
Business & society
Anthropic: compute now eats 58% of revenue, and the company is designing its own chips with Claude's help. According to the updated IPO filing, compute consumed 58% of revenue in the third quarter, up from 48% the previous quarter; the bill was $10 billion, of which $5 billion came from more expensive short-term contracts. The Information reports that the company's 15-person chip design team uses Claude to develop its own AI chips, which it presents to investors as a competitive advantage. Source: TechCrunch, The Information.
ChatGPT: 1.2 billion weekly users, with searches already a fifth of Google's. According to Adam Fry, head of ChatGPT's consumer products, the service added 200 million weekly users in two months, and the volume of queries that trigger a web search is about 20% of Google's search traffic. He did not say where the Google figure comes from. Source: The Verge, Big Technology.
OpenAI: GPT-6 Astra designed a complete hardware device in a week, and an internal model scored even higher on engineering. As a side project, two OpenAI employees had Astra design the enclosure, printed circuit board and software of a device that talks to Codex, without direct human editing, though several rounds of fixes were needed before manufacturing because the model did not always account for manufacturing tolerances. Model use among the company's hardware teams has grown 23-fold since February, aerodynamics and crash-safety research is starting with Mazda, and on the BenchCAD 2.0 benchmark a newer internal model scored 0.875 versus Astra's 0.594. Source: OpenAI, The Decoder.
Acquisitions and investments: Salesforce buys Cresta for $2.7 billion, Fal valued at $16 billion. Salesforce signed a definitive agreement to acquire Cresta, which develops AI agents for phone customer service; according to The Information, the price is $2.7 billion (Missed in our previous roundup: the announcement came out on Monday evening). Fal, which offers AI model inference, raised $1.3 billion at a $16 billion valuation, double its April valuation, and Traversal, which builds site reliability engineering (SRE) agents that find faults in production systems, raised $120 million at a $1 billion valuation. Source: Salesforce, TechCrunch: Fal, Forbes: Traversal.
AI infrastructure: AMD data center revenue up 97%, tech stocks rebound, Meta sells $15 billion in bonds. AMD's third-quarter revenue rose 45% to $13.38 billion, its data center business nearly doubled, and Super Micro's revenue almost quadrupled. After a drop driven by investors worried about AI spending, the Nasdaq rose 0.83% on Tuesday, and Meta received bids worth several times the amount offered for a $15 billion bond sale to fund its AI investments. Source: CNBC: AMD, The Guardian, Bloomberg: Meta.
Data centers and energy: Britain's largest AI data center may be built at the Drax power station, old jet engines put to work in the US. Drax signed a £20 billion deal with Brookfield and Radiant for a data center in North Yorkshire that will eventually exceed 1 gigawatt and could start in early 2028. According to The New York Times, US data centers waiting years for a grid connection are generating power with engines from retired aircraft converted to run on natural gas, while in Russia, Ukrainian drone strikes are pushing plans to move data centers underground, possibly into old mines. Source: The Guardian: Drax, TechCrunch: jet engines, Ars Technica: Russia.
Revolut: switching to self-hosted open models cut its AI coding costs by about 90%. Instead of commercial tools like Claude Code, the British fintech made self-hosted GLM models the default and cut the cost of running coding-assistance AI to roughly a tenth within a year. The company says cheap tokens lead engineers to hand more debugging and small tasks to the models. Source: The Register.
European sovereign AI: Lidl owner Schwarz Digits and France's Bull are building a joint computing service. Under the two companies' preliminary agreement, the AI computing service will operate entirely under European jurisdiction through Schwarz Digits' STACKIT cloud, with servers made by the French state-owned Bull. SAP business software may later be available on it as well. Source: Reuters.
Models that read weights: criticism and a laptop replication follow the GPT-6 Astra experiment. On the experiment covered in our earlier roundup, in which GPT-6 Astra improved a small model's math skills by directly rewriting its weights, spaCy creator Matthew Honnibal wrote that the impressive results rely on a different trick in each experiment, making the overall picture misleading. Meanwhile, Simon Willison showed that Qwen 3.8 27B running on his laptop can handle one of the experiment's weight interpretation tasks. Source: Matthew Honnibal, Simon Willison.
Research: 44% of findings about language models have already "expired," and tracking long texts remains hard. Stanford researchers checked 50 widely cited findings on 20 models: 44% no longer hold, 34% have weakened, and the average lifespan is 15.7 months. Claims about weaknesses go stale fastest, while those about sycophancy and deception persist. On Ai2's new StateBench, however, both GPT-6 Sol and Claude Opus 5.5 fell below 13% on the longest, hardest state-tracking tasks. Source: The Decoder, Ai2 (Hugging Face).
Geoffrey Hinton: we might have been better off without AI, and a major incident could come as soon as this year or next. The Nobel laureate told The Guardian that within a decade we will know whether the risks can be managed, and warned that AI could temporarily cripple the financial or energy system as soon as this year or next. He added that he does "not at all" trust Sam Altman. Source: The Guardian.
Leadership changes: Jill Hazelbaker joins OpenAI, Cat Wu leaves the helm of Claude Code. Uber executive Jill Hazelbaker will become OpenAI's chief corporate affairs officer, reporting to Greg Brockman. Cat Wu, product lead for Claude Code and Cowork, had her last day at Anthropic on Friday, a few weeks after Thariq Shihipar. Source: OpenAI, Cat Wu (X), Business Insider.
Rumors
GPT-6.2: cryptic posts from OpenAI executives and the company's own charts suggest a bigger new model may be close. After Thibault Sottiaux's "Le GPT Fat" post and the "internal model" far stronger than Astra in OpenAI's engineering post, leakers say some users' requests are already being routed to a model called GPT-6.2. Reliability: medium; the internal model's existence is official, its name and release date are not. Source: kimmonismus (X), chetaslua (X), Thibault Sottiaux (X).
Hy4: Tencent's new model is coming soon. The Arena leaderboard announced Tencent's Hunyuan Hy4 model with a "coming soon" label, and Tencent hinted at it with a video captioned "4 Real Productivity." Reliability: high that it is coming, but there are no details or date. Source: Arena (X), Tencent Hunyuan (X).
DeepSeek V5: Chinese media say the base model would be trained on Huawei Ascend chips. Citing supply chain sources, Star Market Daily wrote that DeepSeek would run its next major model on Huawei's Ascend 950 chips even in the pretraining stage, and according to Caiwen, cited by TrendForce, the model could be released in 2027. Reliability: medium; DeepSeek and Huawei have not confirmed it. Source: TrendForce, 21jingji (Star Market Daily).
What this means for you: on Tuesday it became clear from two directions that the rules of AI use are changing fast: the EU Court of Justice's Advocate General says chatbots answering from press articles may infringe publishers' rights, and China promises safety rules covering models' entire lifecycle. If AI tools in your company work with customer data or outside content, consider our AI governance and compliance service, which helps you set out in advance what you can use and how. Revolut's example shows that being tied to a single provider often makes the work more expensive, so with our vendor-neutral AI consulting, we can also work out together which model really pays off where.