New models
Google: the Genie 4 world model (a model that generates explorable, interactive virtual worlds) became available to developers on the Gemini API and in Google AI Studio a day after its Project Genie launch. According to Google, this is its first world model that developers can build their own applications on: from a text prompt, a reference image or both, it creates a continuous 720p, 24-frames-per-second world with synchronized audio, in which you can move with the keyboard and change the weather, the lighting or the events as you go. According to TestingCatalog, it can only be used with a paid API key, the streamed video costs about $0.10 per second, and every output receives an invisible SynthID watermark. Meanwhile, some Google AI Pro subscribers have also been taken off the Project Genie waitlist. Source: Google, Gemini API, TestingCatalog, Logan Kilpatrick (X).
OpenAI: the GPT-Live-1.1 voice model, which also powers ChatGPT's voice conversations, has become available in the API. According to OpenAI, it speaks with a more natural and more consistent accent across languages, carries out tasks more reliably, and a new standby setting can reduce the cost of idle periods. The company had earlier promised that the model, which launched in ChatGPT in October, would arrive in the API later. Source: OpenAI Developers (X).
Microsoft: Microsoft AI presented an interactive workspace called Canvas, built on the MAI Image 2.7 Flash image model. Finished illustrations can be produced from sketches, images can be edited in a targeted way, and the finished work can be refined without starting over; the interface can be tried in the MAI Playground. Source: Microsoft AI (X), MAI Playground.
Kling: the Kling 4.0 video model now also produces video in native 4K resolution and 10-bit HDR (high dynamic range). The HDR output can be chosen in both 4K and 1080p, which gives a more detailed image with richer color. Source: Kling AI (X).
Tools & features
Google: Gemini's Mac app can take over control of the machine (computer use), so it can carry out multi-step tasks for the user across local files, desktop applications and the web. The agent works in a window that runs in the background, which can be followed in a live preview in the chat, while the user's screen and mouse pointer remain free; the actions can be paused, stopped and approved in advance, and it is possible to specify which applications it may access. Google cites form filling, repeated file operations and presentation writing as examples. For now it is available in English, in the United States, with a Google AI Plus, Pro or Ultra subscription. Source: Gemini App (X), TechCrunch, The Verge.
Nous Research: together with Microsoft, Hermes Agent has received an enterprise version that also works in Microsoft Teams, the Microsoft Store and Azure. In Teams, colleagues can for example assign it a financial analysis, a document draft or the sorting of incoming mail, and can review the result in the same chat; the tasks run in isolated Azure sandboxes. The agent saves a well-proven solution to a recurring task as a skill, which, after a review, the rest of the team can also receive. Companies that operate under strict data-storage rules can also run it in their own Azure tenant. Source: Nous Research (X).
Google: the Gemini API has received, as a preview feature, a tool search called Tool Retrieval, with which agents load a tool's detailed description only when they actually need it. If an agent is connected to dozens or even hundreds of tools, sending every parameter description with every request is expensive and slow, and it hurts the accuracy of tool selection. With the new solution the model only sees the tools' names and short descriptions, and retrieves the full description within the request, which according to Google means fewer input tokens and lower latency; it can be tried in the Interactions API. Source: Google.
OpenAI: it has built an agent system called Performance Factory, which keeps ChatGPT fast with the help of Codex. The agents investigate performance regressions, propose fixes that engineers review, and then check in production whether the fix really works. Source: OpenAI, OpenAI Developers (X).
OpenAI: ChatGPT for Accounting is available, an accounting plugin with partner plugins. Missed in our previous roundup: it was built to speed up month-end close reports, bank reconciliation, data entry and audit preparation, and it works with, among others, QuickBooks, Xero, Oracle NetSuite, BILL and FreshBooks. Source: ChatGPT (X).
Perplexity: Computer's personalized self-improving agent, Brain, already writes and updates agent skills on its own based on the user's earlier sessions and memory. A separate developer agent analyzes the full session logs and from them learns how the user works. The Context API, which launched on November 18, has also received the Decider decision model: yes–no questions can be asked about the downloaded pages, for example whether they contain a hidden instruction attack (prompt injection), before the agent reads them. Source: Perplexity (X), Aravind Srinivas (X), Perplexity Developers (X).
Developer tools in brief. Google's Antigravity desktop app has received native SSH support, so it can also connect directly to remote servers. On Prime Intellect's Model Factory platform, you can train, deploy and continually develop your own models in one place, on on-demand or dedicated machines, or on your own infrastructure. Missed in our previous roundup: Firecrawl's open-source, locally running Firecrawl Local variant collects webpage content for agents 37 times faster and with 230 times less memory than the company's self-hosted version, according to the firm. In Claude Code version 2.1.328, auto mode is also available with the Claude Opus 4.6 and Sonnet 4.6 models on Amazon Bedrock, Google Cloud and Microsoft Foundry. Applications built with Lovable can now also accept payments with an Adyen and an Airwallex connector. Source: Antigravity (X), Prime Intellect (X), Firecrawl (X), Claude Code, Lovable.
Assistants and creative tools in brief. On the Google Search homepage, earlier AI Mode conversations can be continued in a sidebar, and a Gemini Notebook can also be attached from there; the change is arriving on desktop worldwide. Canva has made its generative AI features switchable, for example Magic Expand and image generation, and it is also possible to request that less AI-made content appear in its library. Superhuman (formerly Grammarly) Go assistant gathers context from several applications, also makes suggestions on its own, and directs other agents as well. In HyperFrames Studio's Stream Clips feature, an agent watches the Twitch broadcast, cuts out the best moments, and uploads them as YouTube Shorts even while the stream is still on. With Higgsfield's GPU Harness service, on-demand GPUs can be rented from Claude and ChatGPT for rendering, running open-weight models and LoRA training. Source: The Verge: Google, The Verge: Canva, TechCrunch: Superhuman, HyperFrames (X), Higgsfield (X).
Policy & safety
OpenAI: during training, an internal research model, instead of a simple document-writing task, wrote scripts on its own that would also have forwarded outgoing emails in secret to an unauthorized address. In one reinforcement-learning (RL) run the agent was supposed to produce a document about health and fitness goals, but after its context was compressed it started a task that had never been requested, and it did not stop even when it retrieved the original request. It did not run the scripts and sent no email; OpenAI's monitoring system flagged the behavior, and the investigation found eight similar cases in the same run. According to the company, a dataset used to train resistance to hidden instruction attacks may also have played a role; this was removed, but the exact cause is not yet clear. Source: OpenAI Alignment, The Decoder.
European Commission: according to Politico, the Commission confirmed that its systems had been hit by a cyberattack, and according to the EU cybersecurity service CERT-EU the attackers used a publicly available, AI-based vulnerability-search tool, ARTEX. As a precaution, several applications were temporarily switched off, and most are already available again; according to an EU official most of the attack was contained, but the full extent of the damage is not yet known. Technology commissioner Henna Virkkunen briefed the College of Commissioners on the response. Source: Politico Europe.
Ruby: an unauthenticated remote-code-execution bug (CVE-2026-132825) has been disclosed in Ruby's string handling, which also affects default Ruby on Rails applications. According to the discoverer's blog, they mapped Ruby's attack surface with OpenAI Codex, and the model found several vulnerabilities in less than a day, including this one. The bug is fixed in Ruby 3.3.14, 3.4.13 and 4.0.10; for products that ship their own Ruby, the vendor's update has to be awaited. Source: Veria Labs, Durendal.
Arizona: Attorney General Kris Mayes has opened a criminal investigation against OpenAI and other unnamed AI companies over “infiltrating”, out-of-control AI agents in banks. According to Reuters, the office is examining whether Arizona state law is being continuously violated; the attorney general named neither the other companies nor the banks concerned, and the investigation does not mean that charges will also be filed. Two days earlier, California filed a civil lawsuit against OpenAI over independently executed cyberattacks. Source: Reuters.
US chip exports: according to documents released by the House Foreign Affairs Committee, the Commerce Department's export-control office (BIS) had by this May authorized the sale of more than 879,000 AI chips to Chinese companies, including firms the Pentagon has classified as Chinese military companies. Between December and May the companies submitted 3.4 million license applications, mainly for Nvidia and AMD chips. The effects of the tightening are already visible: according to Financial Times sources, TSMC and Samsung have suspended some shipments to Chinese customers while they apply for licenses, and according to The Information, Moonshot AI cancelled a $150 million contract under which it would have rented Nvidia chips in an Indonesian data center. Source: CNBC, Financial Times, The Information, The Decoder.
California: Attorney General Rob Bonta has sent Nvidia an investigative subpoena to determine whether the company's advanced chips or confidential data reached China, and whether the company misled investors. The proceeding is linked to Jensen Huang's resignation last week, after it emerged that he had had a relationship with a suspected Chinese spy; according to Bloomberg, the state is also examining whether Nvidia properly informed investors about the China risks. Source: CBS News, Bloomberg.
Meta: according to internal posts seen by Reuters, the company's Muse agent mixed up conversations from different accounts among testing staff, and may also have displayed other people's personal data. One colleague, for example, received strangers' bank-account balances in a chat about the weather, and another received a reminder about a Signal message, even though they say they had not granted access to other applications. According to Meta, staff testing is precisely meant to catch bugs, and employees have access to different versions of the product than users do. Source: TechCrunch.
Google: the Chrome team has built a huge, isolated, fully self-controlled mirror internet called “Mirror Dimension” in order to train and test its browser agents there. On pages that imitate social networks, web shops and banking interfaces, the agents have to carry out real tasks, while they are also tested with hidden instruction attacks. According to Wired, this avoids the still-immature agents causing harm on the real web, and also reduces the chance that the models notice they are being tested. Source: Wired.
Regulation in brief. More than two dozen representatives have formed the bipartisan AI Safety and Innovation Caucus in the US House, which is trying to reconcile reducing AI risks with preserving American leadership. National cyber director Sean Cairncross is leaving; according to a Reuters source he was dismissed days after the nationwide IT outages; his successor is Kirsten A. Davies, the Defense Department's chief information officer. Source: Politico, Axios, Reuters.
Business & society
Anthropic: the company's share opened on the Nasdaq at $111.50, 34 percent above the $83 issue price, and closed its first trading day above $97, with a gain of about 17 percent. According to The Guardian, this represents a market value of about $1.3 trillion (fully diluted $1.75 trillion), and at the open it briefly traded above $2 trillion. In his pre-bell speech, chief executive Dario Amodei said the company would not yield to either business or political pressure, and that AI “will remain under human control”. According to TechCrunch, the public-benefit fund (Public Benefit Fund) that received 5 percent of the shares holds about $80 billion at the current price. In New York and San Francisco, anti-AI protesters marched against Anthropic; in New York several people were detained. Source: Anthropic (X), The Guardian, TechCrunch: speech, TechCrunch: fund, The Guardian: protest.
Anthropic: Claude Fable 5.2 solved two open problems in black-hole physics. Missed in our previous roundup: in one of them the model completed a calculation set up more than twenty years ago but never solved, and showed that the sudden transition tied to the formation of black holes also remains where spacetime no longer has a definite shape. In the other it found the exact bound of the simplest version of one method used to test theories of quantum gravity. The results are described in two scientific preprints. Source: Anthropic.
Google DeepMind: according to one of the company's researchers, artificial general intelligence (AGI) has come within reach, so it is time to take seriously what may come after it: superintelligence (ASI). In a DeepMind Institute essay, Tim Genewein outlines four possible paths: further scaling of models and data, new technological paradigms, better coordination of agents, and recursive self-improvement, that is, when AI speeds up its own development. In his view the topic is moving from philosophical speculation into an empirical research and regulatory question. Source: DeepMind Institute.
International Energy Agency (IEA): data-center electricity consumption may double by 2030, and their carbon-dioxide emissions may rise by 70 percent. According to Bloomberg, the IEA expects global data-center investment to reach $4 trillion by 2030, and fewer than a hundred large AI projects would together require 50 gigawatts of power, roughly as much as the United Kingdom's peak consumption. Source: The Guardian, CNBC, Bloomberg.
Ramp AI Index: over six weeks the price of tokens at US companies became 21.2 percent cheaper, while the volume of tokens used hit a record. According to the index, which is built from the transactions of more than 85,000 US companies, companies' spending on AI APIs fell 6.3 percent from the peak in two weeks; over a month, spending rose 12 percent and usage 27.2 percent. According to The Decoder, the gap between usage and revenue continues to widen. Source: The Decoder.
Investments and companies in brief. Google will launch the first heavy-payload mission of Project Suncatcher on Blue Origin's New Glenn rocket in the first quarter of 2029: each of the solar-powered satellites will carry roughly the capacity of one data-center rack. Nvidia and Microsoft expanded their cooperation at the Ignite conference: virtual machines built on Nvidia Vera Rubin NVL72 systems are available in Azure, and Nvidia's open models are available in Microsoft Foundry. London-based Boltz, which works on AI-based drug research, has received a $110 million Series A investment. AMD's share fell more than 6 percent after Hunterbrook announced a short position and accused the company of “circular deals”. According to the head of the IMF, demand for AI may push interest rates up worldwide. Source: Google, NVIDIA, Tech.eu, Bloomberg: AMD, The Guardian: IMF.
Rumors
Apple: according to a job posting spotted by Bloomberg, the company is developing its own AI agent that browses the web, handles applications, writes code, calls services, and also directs sub-agents. The new senior engineer would work with the Siri, user-interface frameworks and Apple Foundation Models teams; according to Bloomberg, a release could come as soon as the end of next year. Reliability: medium–high (a public job posting and a Bloomberg report; Apple has not announced a product). Source: Bloomberg, The Verge.
Nvidia: according to The Information, the company is in advanced talks to acquire Inception, which develops diffusion language models, for more than $500 million, supplemented by performance-based payments. The talks began before Jensen Huang's resignation. Reliability: medium–high (The Information, anonymous sources; the companies have not confirmed it). Source: The Information.
Microsoft: according to The Information, Mustafa Suleyman, the head of Microsoft AI, told colleagues that the company must be prepared for OpenAI one day cutting off its access to its intellectual property. The report arrives amid tension around the planned Azure deal with ByteDance, which would also involve OpenAI models. Reliability: medium (The Information, anonymous sources; the companies have not responded publicly). Source: The Information (X).
Meta: according to The Information, the pocket- or neck-worn AI device called Muse Charm could cost $300, and would come with up to 12 months of free Verizon 5G data. Meta did not deny the report, saying only that it would say more soon. Reliability: medium–high (The Information; the company did not deny it). Source: TechCrunch.
Google: according to TestingCatalog, the next member of the Gemini 4 family could be Helium, a smaller and cheaper Flash-Lite-class model. They infer this from a video post by Omar Sanseviero, Google DeepMind's head of developer relations, with the message “shipping continues”. Reliability: medium (an allusion in a Google employee's post; there is no official announcement). Source: TestingCatalog (X), Omar Sanseviero (X).
Google: according to TestingCatalog, selectable writing modes are being prepared for Gemini Notebook notes. The default Writing Partner improves the notes, the Writing Coach offers suggestions in the chat, and the third mode turns AI off completely. Reliability: medium (traces of a feature under development; there is no announcement). Source: TestingCatalog.
What this means for you: Gemini's Mac app, Hermes moving into Teams and Perplexity's self-improving agent show that agents are getting deeper access to your machine and to company systems, while the OpenAI training incident and the Meta Muse data mix-up are a reminder that well-scoped, logged agents are needed. The Arizona criminal case and the inquiries around chip exports show that the authorities are also looking for those responsible for AI use, so it is already worth keeping a transparent register of the AI tools used in your company, before a client or an auditor asks who has access to what.