New models
Google: Gemini 4 Neon has been released, which according to Google delivers frontier-class intelligence with fast, complete task execution in programming, office knowledge work and scientific work. The earlier rumor has thus been borne out: in seven weeks Neon is already the fourth Gemini 4 model, and it sits between Carbon and Hydrogen. It costs $3 per million input tokens ($0.15 for cached input) and $15 per million output tokens, with a one-million-token context and four thinking levels. On Artificial Analysis's composite intelligence index it is third, with 61 points, behind GPT-6.2 Astra and Claude Fable 5.2, at about 17 percent less cost per task than Carbon; on Arena's new agent ranking (the Agent Arena Index) it is second. It is available on the Gemini API and in Google AI Studio, in Gemini for business, in Antigravity and in the Gemini app for paid Pro and Ultra subscribers, where according to 9to5Google it replaces Argon. In GitHub Copilot it can already be selected in the Pro+, Max, Business and Enterprise plans. Source: Google DeepMind, Gemini API, Artificial Analysis, 9to5Google, GitHub.
Google DeepMind: the Genie 4 world model (a model that generates interactive, explorable virtual worlds) has become available in Project Genie for Google AI Ultra subscribers. The world can be modified on the fly with a simple text instruction: the weather can be changed, a new character can be put in, or the plot can take a different direction, and the model now also generates audio. New are an agent that helps build the worlds and a community gallery where worlds can be shared and further reshaped; Pro subscribers can for now join a waiting list. Source: Google DeepMind, The Decoder.
Tencent: the Hy4 flash model, which the company presented a few days ago, has been released with open weights (freely downloadable). In the 224-billion-parameter mixture-of-experts (MoE) model, 12 billion parameters work on each reply, with a one-million-token context, and it also understands images natively. On Hugging Face both the FP8 and the Base versions can be downloaded, and according to the company the larger Hy4 will follow soon. Source: Hugging Face, Tencent Hunyuan (X).
Nvidia: the open Nemotron-Labs-Sigma models generate tokens not one by one, but in blocks, in parallel. According to Nvidia, this makes text generation faster, and makes it easier for the model to try several solutions at once to a mathematical or programming task. The method is described in a scientific paper, and the models can be tried out. Source: NVIDIA AI (X), arXiv.
Pika: as a research preview it has presented Enact, with which the user can, as in a film shoot, "direct" the AI-generated characters during the scene. Enact is built on the Pika H3 Flex model, and according to the company it aims to bring video-making closer to a shooting situation: the scene responds to the director's instructions. Source: Pika (X).
Tools & features
Google: under the name Stitch Loop, Google Labs has presented a long-running agent that, watching the live product's data, independently proposes and carries out improvements. Stitch Loop connects to GitHub and to the product's data (logs, user behavior, feedback) through more than 180 integrations, its own MCP servers (standard connectors that tie agents to external tools) or the command line. In every round it makes a proposal backed by evidence, builds the chosen fix in an isolated cloud environment, and opens a pull request, which a human approves. Built-in rounds deal with security (with the help of Gemini 4 Carbon and the CodeMender security agent), response time and accessibility. The waiting list is open to everyone, and Ultra subscribers get priority. Source: Google, TestingCatalog.
Mistral: the Vibe desktop app has been released for macOS and Windows, in which several tasks can be given to the agent at the same time. Missed in our previous roundup: the agents can run on the local machine or in the cloud, they work side by side, and their changes and pull requests can be reviewed in the same window; parallel tasks are separated by Git worktrees (separate working directories). Work mode already also has access to local files. A Linux version will come later. Source: Mistral AI (X), The Decoder, TestingCatalog.
Perplexity: the Context API has launched, which reads in at most ten web pages per call and returns clean Markdown or HTML content, links and excerpts from them. According to the company, it returns usable content for 92.3 percent of URLs, costs $1 per thousand URLs read, and can be used in any agent framework. Perplexity's Teammates and Multiagent features are now also available in the iOS and Android apps for Max subscribers, with voice control for Teammates. Source: Perplexity, Perplexity Devs (X), Perplexity (X).
GitHub: eight further services, including Supabase, Box and Calendly, can be connected to Copilot through MCP servers reviewed by GitHub. The connectors are available in public preview in every Copilot plan, and join the more than 25 services already there (for example Slack, Figma, Sentry). Team-level customizations have also become generally available: team maintainers can set approved agent extensions as the default, and at GitHub Team organizations free security risk assessments run automatically to find leaked keys and bugs in the code. Source: GitHub: connectors, GitHub: team settings, GitHub: risk assessment.
Snapchat: with the new Easy Lens CLI, developers can build augmented-reality (AR) lenses with the AI coding assistants they already use, for example Claude Code, Cursor or Codex. The tool is built on the web and mobile Easy Lens, which makes a lens from a single instruction; the command-line version leaves the technical work to the assistant. Source: TechCrunch.
Assistants in brief. On the Google Search AI Mode interface, temporary conversations can be started, which do not go into the history and do not affect personalization. Thanks to a redesign, the Gemini app also runs on lower-memory Android (Go edition) phones. ChatGPT's dictation has become faster and more reliable, and the text of the speech now appears while you are speaking. Source: Google: AI Mode, Google: Gemini, Greg Brockman (X).
Creative tools and hardware in brief. Suno's Visuals feature produces images and 15-second videos that automatically fit the generated songs. Higgsfield MCP 2.0 makes the company's full creative toolkit available inside Claude, with workflows of several agents. Compact Nvidia RTX Spark desktops from ASUS, Dell and MSI have appeared, with up to 128 GB of unified memory, for running local AI agents and models. Source: The Verge: Suno, Higgsfield (X), NVIDIA.
Policy & safety
Nginx: two critical, unauthenticated remote-code-execution bugs (CVE-2026-123633 and CVE-2026-123007) have been disclosed in the world's most widely used open-source web server. The bugs were found independently by Wiz, Ikotas Labs, Metnir, Rapid7 and Xint; according to Forbes, Xint used OpenAI's GPT-6.1 Sol and GPT-6 Luna models. Both bugs have a severity of 9.2 under CVSS 4.0, can only be exploited with HTTP/2 turned on, and one of them affects the software all the way back to version 1.9.5 from 2015. Operators should update to the patched version. Source: Wiz, Forbes.
Microsoft: attackers uploaded two malicious versions (1.26.1 and 1.26.2) to the official PyPI package of the company's agent-development toolkit, Agent Framework. The versions published on November 17 stole cloud access credentials and other sensitive information, and ran as soon as Python started, even if the package was not imported. Users of the package, which is downloaded about 200,000 times a month, should check the installed version and rotate the keys stored on the affected machines. Source: BleepingComputer.
PromptArmor: under the name FrontierBreach it has released a benchmark of how often workplace AI agents can be misled by an instruction hidden in an email, a document or a skill file. In prompt-injection (hidden-instruction) attacks the agent may send out data or run the attacker's code. According to the company, existing tests underestimate this risk; in the measurement every frontier model tested was successfully deceived by at least some of the attacks, and the differences among the models were large. Source: PromptArmor (X), TestingCatalog (X).
OpenAI: it has set out in detail how it handles the risks of frontier models, from research through internal use to public release. The company builds on four layers of defense: model alignment, isolation, defenses that block harmful requests, and continuous monitoring. Since the July Hugging Face incident it trains frontier models without live internet access, monitors every internal deployment that uses tools, and holds a safety review before every stage. According to OpenAI, this is why it releases models more slowly; it held back GPT-6.1 Astra, for example. Source: OpenAI.
Anthropic: before the IPO, the Center for AI Safety and an unusually mixed coalition (safety researchers, advocates of open source, AI ethicists and supporters of speeding up AI development) signed the so-called MisAnthropic statement. According to the statement, there is a contradiction between Anthropic's "safe" image and the fact that the company wants to build self-improving (recursively self-improving) AI and to concentrate power. According to Dan Hendrycks, the director of CAIS, a company that is weakening its earlier safety commitments cannot credibly call itself safe. Source: Center for AI Safety.
United States: intelligence and homeland-security leaders told Congress there is no sign that the nationwide IT outages of recent days are connected, or that a cyberattack caused them. In a letter, Senators Maria Cantwell and Mike Rounds asked Jay Clayton, the director of national intelligence, and the White House official responsible for AI, to find and have fixed the vulnerabilities in critical infrastructure, because AI finds and exploits software bugs faster than organizations can patch them. According to the officials, the disruptions were caused by software patches installed to handle AI-related risks; how large a role AI played in this is not yet clear. Source: The Guardian.
California: under proposals from Governor Gavin Newsom, the state would prescribe mandatory baseline safety rules for every AI developer, and would gain the power to restrict the development and use of self-improving (recursively self-improving) AI. The proposal would require independent, external safety testing and auditing at every stage of the models' training, development and release, would broaden the reporting of safety incidents, and in an emergency the state could order a model to be shut down at once. According to Politico, however, it would not make it mandatory to develop a separate emergency shut-off (a kill switch). Source: Governor of California, Politico.
Banks: Citibank has amended its terms of use, under which anyone who gives a third-party AI agent access to their bank account is themselves responsible for the operations carried out through the agent, even if it is compromised. According to TechCrunch, JPMorgan and Bank of America earlier announced a similar step: the banks will not undertake to reverse the agents' deliberate or accidental transactions. Source: TechCrunch.
Business & society
Anthropic: it priced its IPO at $83 a share, above the $70–78 range, which according to CNBC, at $81.3 billion, is the largest offering of all time. The pricing above the range in earlier reports has thus been borne out. In the upsized offering, 980 million shares are being sold, of which 956 million come from Anthropic itself; trading starts on November 19 on the Nasdaq under the ticker ANT. According to CNBC, the fully diluted market value is about $1.5 trillion, and the offering surpasses SpaceX's record from June. Epoch AI highlighted that, according to the prospectus, the company spent $10 billion on compute capacity and infrastructure in the third quarter alone. Source: Anthropic, CNBC, The Guardian, Epoch AI (X).
OpenAI: together with AMD, its models achieved in less than five months improvements of as much as 14 percent in performance, power and area (PPA) in AMD's flagship GPU designs. According to the company, this is roughly equivalent to a year of progress, and AMD has already built some of the changes into a design sent to manufacturing; on certain blocks of the next-generation flagship processor, single-digit improvements have been confirmed. On OpenAI's own inference chip, Jalapeño, the models' suggestions reduced the area of the vector and matrix units by 13 and 10 percent. Together with Synopsys, the company is training a model tailored to chip-design software, named GPT-Synopsys. Source: OpenAI, The Decoder.
OpenAI: according to an internal measurement, after 23 percent in March, in October agents already completed 46 percent of two-to-four-hour research tasks on their own. For four-to-eight-hour tasks the share rose from 13 to 35 percent, but the differences by type of task are large. A researcher's daily agent use, at list prices, had a median of $2,400 in October. Researcher time spent on safety and alignment rose 77 percent from January to September; according to the company, labs should be required to report signs of AI self-improvement. Source: OpenAI.
Nvidia: at the first all-hands meeting without Jensen Huang, interim CEO Colette Kress promised that staff would get a pay rise and more stock in December, months earlier than planned. According to Wired, of the 700 questions submitted in advance by employees, the largest number were about pay. Senator Tim Scott announced that Huang has agreed to appear also before the Senate Banking Committee, and according to CNBC some of the employees still want Huang to return. Source: Wired, CNBC: Senate, CNBC: employees.
Palo Alto Networks: it is buying Applied Compute, which helps with the enterprise customization of open-weight AI models, for $4.5 billion in stock. According to The Wall Street Journal, the startup was still valued at $3.25 billion this month; on the same day the cybersecurity firm also reported stronger-than-expected quarterly revenue. Source: Reuters, The Wall Street Journal.
KPMG: the firm's UK unit ran a parallel audit of an FTSE 250 company with more than a hundred AI agents, and in some cases the agents dug deeper and asked better questions than the human team. In the experiment, named Project Exponential, the agents worked alongside the traditional team across the full audit process, not only on individual administrative sub-tasks. Source: Financial Times, European Business Magazine.
Society in brief. According to a Gallup survey in 16 countries, more than a third of 15-to-22-year-olds rely on AI for company, among daily users the share is 50 percent, and they are more likely to be affected by loneliness. At the ARIA awards, Elton John stood up for Australian musicians who are fighting the use of their works to train AI. According to The New York Times, school districts in New York, Los Angeles and Chicago are looking for an alternative to Google Search, after the company did not offer a version without AI features for schools. Source: Wired, The Guardian, The New York Times.
Investments and companies in brief. Intel has signed its first external customer for its advanced 18A-P manufacturing technology. Tencent's third-quarter operating profit fell for the first time in years, while its AI investments grew fivefold, to about $12 billion. Masayoshi Son, the head of SoftBank, called it a mistake that they did not invest in Anthropic in 2023. According to the European Central Bank, the borrowing boom tied to AI shows a historical parallel with the subprime mortgage bubble. OpenAI has entered a partnership with the publisher of Le Figaro, so the paper's articles appear in ChatGPT, and Paris-based Remanence has received a $30 million seed investment for training and fine-tuning open-weight models. Source: CNBC: Intel, Bloomberg: Tencent, Financial Times: SoftBank, Financial Times: EKB, OpenAI: Le Figaro, Axios: Remanence.
Rumors
Trump administration: according to Reuters sources, the administration is considering requiring US companies to indicate on their websites if they use Chinese AI models. The aim would be to curb the use of Chinese models. Reliability: medium–high (Reuters, anonymous sources; there is not yet a decision). Source: Reuters.
Sierra: according to The Information, Bret Taylor's company, which builds customer-service agents, would raise about $1.5 billion at a valuation of about $30 billion, almost double the $15.8 billion of May. In October the company reported an annual revenue run-rate of more than $400 million. Reliability: medium–high (The Information, anonymous sources; the company has not confirmed it). Source: The Information (X).
Microsoft: according to The Information, the company is close to selling AI services to ByteDance worth about $500 million, mainly access to OpenAI models through Azure. According to the paper, this is a further source of tension between Microsoft and OpenAI, because OpenAI objects to the deal. Reliability: medium–high (The Information, anonymous sources; the companies have not confirmed it). Source: The Information (X).
Telegram: according to TestingCatalog, the company is building its own AI agent named Dictator for Telegram on macOS, with its own Linux virtual machine. Based on traces seen in the beta versions, the agent may be built on the Nous Research Hermes Agent framework, would support several models (ChatGPT, Claude, OpenRouter, local models) and plugins, with per-operation permissions, and would not have access to personal files. Reliability: medium (signs seen in beta code and in a promotional video; Telegram has not announced anything). Source: TestingCatalog.
Leopold Aschenbrenner: according to Financial Times sources, the investor would restart his AI-focused fund under the name "Situational Awareness 2.0," with debt and outside capital, four months after Citadel rescued it. Reliability: medium–high (Financial Times, anonymous sources; there is no official announcement). Source: Financial Times.
Upcoming Google models: according to TestingCatalog, after Neon the Gemini Helix and the Nano Banana 3 Pro image model built on Gemini 4 may arrive soon, and a model named "Cobalt" has also appeared in the code of Google AI Studio. Reliability: low–medium (a forecast by a leak-watching site and traces in the code; Google has not announced a date). Source: TestingCatalog (X), TestingCatalog (X): Cobalt.
What this means for you: with the arrival of Gemini 4 Neon, prices among frontier models have been reshuffled again, so it is worth comparing, task by task, which model delivers the quality you need at the lowest cost, and building the system vendor-independently, so that you can switch quickly to a better offer. The compromised Agent Framework package, FrontierBreach, which measures hidden-instruction attacks, and the banks' new terms also show that the security of agents is your responsibility too, so only let vetted, tightly permissioned and logged agents into your mail, your code or your finances.