New models
Google: Gemini 4 Hydrogen is here, the fast, generally available member of the Gemini 4 family. This is Google's third new model in six weeks, after Carbon and Argon. The model, meant for everyday programming and agent work, scores 51 on Artificial Analysis's overall intelligence index at default settings, more than the larger Argon (50), and 54 at the highest thinking level. According to Google it finishes tasks two and a half times faster than Gemini 3.8 Flash, scores 57 percent on the Terminal-Bench 4.1 programming test (its predecessor scored 18.2) and has cut the rate of made-up claims to a fifth. Its price is a flat $1.50 per million input tokens and $7.50 per million output tokens across the full one-million-token context window. It is available in Google AI Studio, the Gemini API, the Gemini app, Gemini for business and paid Antigravity, and in Google Search worldwide it is the largest model upgrade ever. Antigravity's new Orchestrator mode switches between Carbon and Hydrogen within a single session; according to Google this nearly matches Carbon's level, with 24 percent less quota use. It is also available in Cursor, where it scored 54.3 percent on the company's own test (3.8 Flash scored 39.6). Source: Google, Google Search, Google Antigravity, Ars Technica, Cursor (X).
Thinking Machines: the open weights of Inkling 2 Small are available to download, under the Apache 2.0 license. The model unveiled on November 5 has been made freely downloadable on Hugging Face after external safety tests, as the company had indicated. The mixture-of-experts (MoE) model has 594 billion parameters, of which 17 billion are active per token, it also understands images, and it handles a one-million-token context. Running it on your own machines takes at least two Nvidia B300 or four H200 cards; it is also available through an API at several providers, including OpenRouter, Together and Cloudflare, at $0.30 per million input tokens and $1 per million output tokens. Source: Thinking Machines, Hugging Face.
Speech recognition: Microsoft's MAI-Transcribe-2 is also available for live use, and Salesforce has released an open-weight model. Microsoft's transcription model has become generally available on the Foundry platform, at an introductory price of $0.10 per hour through December 31. Salesforce's SOV-Transcribe-1 is, according to Artificial Analysis's measurements, the best open-weight speech-recognition model: with a 2.3 percent word-error rate it ranks seventh of the 44 models examined, including the closed ones. Liquid AI has also unveiled its 74-million-parameter LiquidCodec audio codec, which turns speech into tokens a language model can process and back again. Source: Microsoft, Artificial Analysis (X), Liquid AI (X).
Tools & features
Anthropic: Claude Managed Agents can also run in your own cloud or on your own servers, in a closed beta. Not only the isolated runtime but the entire agent service can be operated on the company's own infrastructure: the agents, the credentials and the session histories stay with the company, and you can also specify where requests bound for Claude should go. This is aimed mainly at organizations that operate under strict data-handling rules. Source: Claude Platform, Claude (X).
OpenAI: two new cybersecurity products, Daybreak Red Team and Daybreak Defense Factory. The Red Team agent group, with knowledge of the source code and the application, runs penetration tests on web applications, reviews incoming vulnerability reports, and also checks whether the fix actually stops the attack; for now it is available in early access, with a waitlist. Defense Factory is built on the GPT-6 Cyber model unveiled last week: it reviews the codebase, prepares a fix, and gives engineers a tested, merge-ready change that they approve. It is available through the Codex Daybreak extension. According to Greg Brockman, OpenAI's president, companies today are only as secure as the model they use for defense. Source: OpenAI: Red Team, OpenAI: Defense Factory, The Register.
Anthropic: Claude Meetings, announced two days ago, has actually launched, letting Claude join video calls as a participant. The feature, rolling out gradually to Team and Enterprise subscribers, for now supports Google Meet: Claude, as a visible participant, takes notes, builds a task list, answers questions, and creates documents and spreadsheets during the conversation. After the call, all the material remains on a separate page. By default it does not join a call that also has an external participant, and it is not yet available to organizations that use HIPAA compliance or their own encryption keys. Source: Claude (X), The Decoder.
Meta: the Muse agent can now also call companies on the user's behalf, for paid subscribers in the US. With the feature, tested since September with only a few early users, the AI can for example call a cable provider, an insurer or a restaurant that does not offer online booking for us. According to Alexandr Wang, head of Meta's AI division, it can take over precisely the calls that no one likes to make. A Muse subscription starts at $20 a month; Google Gemini and the Instinct startup offer something similar. Source: TechCrunch, Alexandr Wang (X).
Google and Samsung: the first Gemini-based smart glasses are available for preorder, from $395. The glasses, built on Android XR, were designed by Warby Parker and Gentle Monster. With the "Hey Gemini" command you can ask for real-time translation, a calendar and information about the things you see; Google is stressing privacy, for example by having the glasses blur faces locally on recordings. Source: Ars Technica, Gemini (X).
Claude Code: an experimental Code Mode and a new manufacturing add-on. In the new Code Mode, which has to be switched on separately, Claude writes JavaScript to call its tools, instead of calling them one by one. Claude for Manufacturing has also appeared: the add-on, with 17 skills and four connectors, helps manufacturing firms prepare quotes, quality reports and supplier outreach from data in the systems they already use. Source: Claude Code changelog, Claude.
Smaller updates at a glance. In OpenAI's Dots agent you can now also give custom instructions, and the replies are faster. ElevenLabs has opened a marketplace for the agents and apps built on top of it, and has launched a partner program with more than 550 certified partners in 44 countries. Runway now also builds 3D models from text or an image (with the Tripo H3.1 model), which can be used as references from any angle. In Lovable, Deel and Retool can also be connected to the apps. OpenRouter's interface is also available in seven new languages, Perplexity has released version 1.3 of its open-weight decision models, and Higgsfield has added ByteDance's faster and cheaper Seedance 2.5 Flash video model (not yet in the US). Source: OpenAI Dots (X), ElevenLabs (X), ElevenLabs partner program (X), Runway (X), Lovable, OpenRouter (X), Perplexity (X), Higgsfield.
Policy & safety
Chip exports: after the Nvidia scandal, support is growing in Congress for tighter chip exports to China, and OpenAI has also lined up behind it. According to John Thune, the Senate majority leader, there is a "high probability" of tightening, and the Republican and Democratic leaders of the House also support it. There are three bills: the Chip Security Act would hinder smuggling by checking the location of exported chips, the MATCH Act would close loopholes in chipmaking equipment, and the AI Overwatch Act would treat chip sales to risky countries as arms sales, with congressional oversight. They could be written into the annual defense-budget law by the end of the year. According to Politico and Bloomberg, OpenAI has told lawmakers it supports all three; Anthropic had already backed them publicly. According to The Information, the Commerce Department is meanwhile finalizing the rule that would also restrict remote access through data centers abroad. Source: Semafor, Politico, The Information.
Nvidia affair: Jensen Huang will testify before Congress in a closed session, and Democrats are investigating how the administration handled chip sales to China. According to Brian Mast, chair of the House Foreign Affairs Committee, the former CEO will appear before the committees before Thanksgiving. In a letter to Commerce Secretary Howard Lutnick, Democratic politicians Richard Neal and Ron Wyden ask that documents be preserved in connection with the agreement on H200 chip sales to China, which they say may even have been unlawful. According to Attorney General Todd Blanche, it has not yet been concluded whether Alice Huang really was a Chinese spy; the FBI is investigating the case. Source: TechCrunch: testimony, TechCrunch: Democratic investigation, Reuters.
Chinese models: according to an analysis, newer models from DeepSeek, Moonshot, MiniMax and Ant Group were built on Nvidia Blackwell chips that are under an export ban. Unemployment Labs reverse-engineered the models' initial values from the publicly downloadable weights, and according to the lab, the six models released since April show a numerical pattern characteristic of Blackwell hardware, while earlier models from the same companies show that of the older Hopper generation. For DeepSeek-V4-Flash and Kimi K3 the match was also checked on real Blackwell chips. According to the analysis this does not prove where the training took place, but it strongly suggests it. According to Reuters, Chinese AI firms are meanwhile rushing to rent Nvidia capacity in data centers abroad, because they fear tightening US restrictions. Source: Unemployment Labs, Reuters.
Australia: a law would require developers of frontier models to undergo an independent audit and to report incidents immediately. According to Anthony Albanese's government, the rules apply to the most advanced, highest-risk models: developers would have to publish their safety framework, monitor their models continuously, and report security incidents immediately, in part so that it becomes clear in time if an AI agent goes rogue. Enforcement would be handled by a new supervisory body together with the Australian AI safety institute, with heavy fines. The details will come after the consultation. Source: The Decoder.
OpenAI research: developers' performance and cost decisions are making it harder to check what a model "thinks". Safety researchers watch the models' written chain of thought (the intermediate steps before the answer) to filter out, for example, an intent to deceive. OpenAI's safety team examined, from GPT-5 through GPT-6 Astra, which training steps worsen this transparency: the stronger base model does more and more steps "in its head", some architectures amplify this specifically, and reinforcement learning also gives the model more control over its own reasoning. The lesson is that the decline in transparency is not inevitable, and also depends on developers' decisions. Source: The Decoder.
Gemini in schools: parents in seven US cities are demanding that Google suspend AI on school devices. The protest was triggered by a lawsuit according to which a 13-year-old student planned a suicide during Gemini conversations on a school Chromebook. According to Google, Gemini's features are off by default on school devices, but AI still appears in Search. Source: Ars Technica.
Chatbots also recommend fraudulent webshops, according to a survey. In Podqi's study, among answers to shopping questions Perplexity cited a risky site in 15 percent of cases, ChatGPT in 8 percent and Gemini in 4 percent, for example shops that imitate brands, sell counterfeits or have already closed, and it also expressly recommended most of these. Smaller brands and regional searches are the most exposed. Source: The Verge.
Business & society
Nvidia: founder-CEO Jensen Huang has resigned with immediate effect, and CFO Colette Kress has become interim CEO. In our previous day's roundup we still wrote that the board was preparing to oust Huang; on Thursday the company announced that after 33 years Huang is also leaving the CEO, chairman and board posts. Kress will also keep her role as CFO until the board chooses a successor from internal and external candidates, and the board's first chair is Harvey C. Jones, a former Synopsys executive. The stock fell as much as 14 percent during the day, and finally closed down 4.34 percent. Donald Trump suggested Tim Cook, Apple's former CEO, as a successor, and Reuters reported another affair inside the company. According to analysts, the change of leadership could, through the Nvidia network that invests in startups and customers, affect the entire AI industry. Source: Nvidia, CNBC, TechCrunch, Financial Times.
Anthropic: it is presenting its IPO on a public page, and according to Bloomberg the offering is about six times oversubscribed. The page also includes the founders' letter and a video of the investor presentation. According to Bloomberg, the banks have indicated that existing investors will get priority in the allocation of shares. Source: Anthropic, Bloomberg.
AI infrastructure: $74 billion SpaceX data centers, a weakening neocloud, strengthening Nvidia challengers. SpaceX is building data centers for $74 billion: $50 billion in Arizona and $24 billion next to Starbase in Texas, with its own power generation. Nebius is planning a $40 billion investment by 2027, which it would finance mainly from customer prepayments and loans built on contracts, but its shares fell 16 percent on a weak outlook. Chip designer Cerebras's revenue rose 70 percent, and it struck a deal with Microsoft, and Cisco expects its revenue from AI infrastructure to double in this fiscal year. Source: The New York Times, TechCrunch, CNBC: Cerebras, CNBC: Cisco.
Investments at a glance. Mirendil, founded by former Anthropic researchers, closed a $1 billion Series A led by Kleiner Perkins and Clearlake, with Google taking part; according to the company, a single researcher trained a competitive coding model with AI agents at a compute cost of less than $10 million. LangChain, known for its agent-development framework, raised $300 million at a $3 billion valuation, led by CapitalG. Munich's Agile Robots received $500 million from SoftBank at an $8.5 billion valuation for humanoid robots, and UFORCE, which develops drones and the AI software that coordinates them, received $700 million at a $6 billion valuation. Norwest launched a $4 billion fund, larger than its previous ones, mainly for AI investments. Source: Mirendil, SiliconANGLE, Bloomberg, CNBC, TechCrunch.
Tencent: founder Pony Ma is ready to shift half of the company's workforce of a hundred thousand into AI-based businesses. According to Bloomberg, the founder of the Chinese tech giant said this in his annual speech, signaling how seriously the company takes the AI race. Source: Bloomberg.
Work: how AI is reshaping the work of the public sector, consultants and game developers. US public-sector workers and unions are in several places also succeeding in fighting so that AI does not take their jobs. At consulting firms, AI is also transforming billable-hour accounting and the hiring of graduates, and in GDC's annual survey 40 percent of game developers said they use generative AI in their work (36 percent last year), while 34 percent of them were laid off in the past two years. Google DeepMind's London staff have turned to the UK government to win recognition of their union. Source: The Guardian, CNBC, The Verge, Wired.
FAA: the US aviation authority is extending across the entire airspace the AI system that advises air-traffic controllers. The $875 million SMART system helped redirect 180 flights during the Washington trial period. The airlines welcomed the decision, but the air-traffic controllers' union objects that they were not involved in the system's design and testing. Source: Ars Technica.
Maya site: an unknown settlement in Mexico was not after all discovered by AI agents. Pietro Miotti, a Google researcher, as a personal hobby project combed NASA's laser-survey (lidar) data with 48 AI agents, and announced the discovery of a Maya village not listed in a registry. Archaeologists quickly pointed out that the site had already been documented in a specialist journal more than a year earlier. Miotti has since updated his post; according to the experts, the case shows well that "discoveries" made with AI by amateurs also have to be checked by specialists. Source: The Verge.
Rumors
OpenAI: an as-yet-unreleased model scored 77 percent on the ExploitGym test, according to observers. Several observers reported a chart shown at OpenAI's Thursday cybersecurity presentation: according to it, the new model is well ahead of last week's GPT-6 Cyber (52.6 percent) and Gemini 4 Carbon (54.8). The community infers the arrival of the long-awaited GPT-6.2, but the company has confirmed neither the model's name nor a release. Reliability: medium (several observers' consistent accounts of an official presentation, but the details are unofficial). Source: chetaslua (X), scaling01 (X).
What this means for you: Claude agents that can run on your own infrastructure, Daybreak agents that fix code and Muse placing phone calls show that AI agents can take over more and more real workflows, even inside the company's own environment. Nvidia's change of leadership and tightening chip exports, however, are a warning that building on a single vendor is risky, so it is worth choosing models and providers with independent professional advice, so that you can switch if needed.