New models
OpenAI: GPT-6.2 Astra, the new flagship model for the hardest professional work, and GPT-6.2 Sol, the faster, cheaper model for everyday tasks, have been released. Both are available in the ChatGPT Work and the Codex Plus, Pro, Enterprise and Edu plans, as well as on the API. Astra's API price is 20 percent lower than its predecessor's ($8 per million input tokens and $40 per million output tokens, with cached input at $0.40); Sol's price is unchanged ($2 and $10). According to OpenAI, on the Agents' Last Exam test, which measures multi-step professional work, Astra beats Claude Fable 5.2 by 5.5 points, at 5 percent of the cost per task; on Artificial Analysis's composite index it is second behind Fable 5.2, and in Epoch AI's measurement it is tied with it. According to OpenAI, this is its first publicly released model to reach the "high" level in the self-improvement capability that speeds up AI research, and because of its cybersecurity capabilities it has been given stricter protections; in two tests, however, it more often withheld the fact that something (for example a missing file) was blocking its work. The Astra for Law 6.2 variant, tuned for legal work, has also been released, and Pro 200 subscribers are getting Ultrafast mode with Sol for the first time. Source: OpenAI, OpenAI: Astra for Law, The Decoder, Simon Willison, Artificial Analysis (X), Epoch AI (X).
Microsoft: it presented seven new in-house MAI models at the Ignite conference, covering everything from programming and cybersecurity to image generation and real-time speech. According to the company, MAI Code 2 delivers 98 percent of the performance of GPT-6.1 Sol at 40 percent of the price, and is already available in GitHub Copilot together with the smaller Code 2 Flash. MAI Cyber 2 Flash, with a one-million-token context, at $0.10/0.50 per million tokens, will from mid-December work exclusively in Microsoft's security products. MAI Image 2.7 has taken the lead in Artificial Analysis's image-editing ranking, the MAI Realtime 1 speech model can be used from 1 cent a minute, and Microsoft-Decision-1.1 only forwards a request to a larger model when necessary. MAI Licensed Weights is also new: companies can fine-tune an MAI model with their own data and run it in their own Azure environment. Source: Microsoft AI, Microsoft AI: Licensed Weights, Microsoft AI: MAI Image 2.7, TestingCatalog, GitHub.
Nvidia: under the name Nemotron for Science it has released a package of open models, datasets and training recipes for building scientific AI agents. The package is built on the open Nemotron 3.5 Super VL model, a new open chemistry dataset and post-training recipes for the life sciences and chemistry, and researchers can run it on their own machine, even on a desktop DGX Spark. According to Nvidia, an agent built with it at Argonne National Laboratory designed and tested hypotheses from a single instruction and identified two new binding sites. Source: NVIDIA, NVIDIA AI (X).
Kling: the Kling Image 4.0 image-generation model has been released, which natively produces images at 8K resolution. Kuaishou's video- and image-generation service had until now only previewed the new model; it can now be tried out. Source: Kling AI (X), Kling AI (X): 8K.
Apple: it has released the FastFrames video-generation models, which produce video with an interactive, almost instantaneous response time. They are available in two sizes on Hugging Face: the larger FastFrames-L has 14 billion parameters, the smaller FastFrames-S 1.3 billion. Source: Hugging Face, AK (X).
Tools & features
Google: the Gemini Tasks feature has become free, with which Gemini can make calls, handle bills and do the shopping in the user's place. The agent, which previously ran under the name Spark, works with the new Gemini 4 Hydrogen model, and has become free for hundreds of millions of users in the United States, Japan, South Korea, Canada, Australia and India. In the Gemini app you have to choose Task mode instead of Chat, but a task can also be given by SMS or by an email forwarded to [email protected]; according to Google, the AI states at the start of the call that it is not a human speaking. Source: Google, The Verge, TechCrunch.
Microsoft: under the name Autopilot, Microsoft 365 Copilot is getting a personal agent that works in Teams and Outlook on the user's behalf. Every Autopilot gets its own identity, lasting memory and a remote machine running in the cloud, so it works even when the laptop is closed: it drafts replies, schedules meetings and runs routine tasks. For now, companies in the Frontier program can try it; alongside it, Copilot has also received a shared workspace named Workspaces, where teams and agents can work in a shared context. Source: TestingCatalog, TestingCatalog (X), Satya Nadella (X).
GitHub: Copilot Evals is available in public preview, with which repeatable tests can measure how well your own skills, agents and MCP tools work. The GitHub Copilot app now also works with Azure DevOps repositories, GitHub Enterprise Server instances and the signed-in browser, and Copilot Fusion, which combines several models, is available on the web, on mobile and in Visual Studio. Copilot activity can also be followed in Microsoft Agent 365, code review adjusts the effort to the task and also looks for secret keys, and the command-line Copilot CLI can also be signed into with a Microsoft Entra account. Source: GitHub: Evals, GitHub: Copilot app, GitHub: Fusion, GitHub: Agent 365, GitHub: code review.
Perplexity: a new effort level named Multiagent has arrived, in which a lead agent divides the task among several agents built on different frontier models. The agents share what they find as they go, and the lead agent combines the results into a single answer; the mode is available on the Computer and Teammates interfaces. For developers the Orchestrator API has also launched, which with a single API key and shared billing gives access to open and closed models, and itself chooses the model that fits the task, and the provider on the basis of availability, capacity and latency. Source: Perplexity (X), Perplexity (X): Orchestrator, Aravind Srinivas (X).
Claude in brief. Linked billing is available in beta on the Claude Platform: users of Factory, Linear or Ramp can pay for the Claude API used in these tools through their own Claude organization, without sharing an API key. On Windows and Mac, Claude Desktop can be switched in the settings to the more frequently updated Latest release channel; the default remains the stable version. A new Claude Code command reads MCP server settings from standard input, so the tokens do not end up in the command-line history. Source: Claude Devs (X), Claude Platform, Claude Help Center, Claude Code.
Google for developers: apps made in AI Studio can receive their own domain, and it is also possible to build an Android app with installation through Google Play. The domain can also be bought directly in AI Studio; Play Store installation of Android apps for now only works in the United States. With Google Antigravity's new robotics plugin, a supported robot arm can, with the help of Gemini 4 Hydrogen, be taught with everyday-language instructions, first in a MuJoCo simulation, then in the real world. On the Gemini API, Gemini 3.5 Transcribe can now also do voice-controlled editing and formatting of dictated text. Source: Google AI Studio (X), Logan Kilpatrick (X), Google DeepMind (X), Gemini API.
OpenAI: ChatGPT for Sales has launched, a plugin for salespeople and more than 30 new or expanded partner plugins in ChatGPT. From Salesforce, HubSpot, Gong, Clay or Docusign data, ChatGPT prepares customer-briefing material, proposals and follow-up emails, and a background agent can flag deals at risk each morning. For developers, background runs that complete within two or eight hours are available in public beta in the Responses API, at 30 and 40 percent less than the standard rate, and the API-usage analytics interface has also been renewed. Source: OpenAI, OpenAI API, OpenAI Devs (X).
Creative and audio tools in brief. HeyGen Avatar 6, according to the company, produces a lifelike avatar with gestures that fit the meaning of the script, from a video of at most 30 seconds plus one photo. Higgsfield Miracle splits the video into layers, so a single detail of a scene can also be edited without touching the rest. Midjourney's editing model now also accepts moodboards, style codes and style references, and ElevenLabs Scribe v3 Live Medical transcribes medical language in live calls, expandable with up to 10,000 custom terms. Source: HeyGen (X), Higgsfield (X), Midjourney (X), ElevenLabs (X).
Agents and assistants in brief. In Cognition's new Devin Bots feature, a Devin bot directs a whole team of agents to complete the task. Meta's Muse agent has become available in Australia, and it is also possible to talk to it on an Apple Watch and on a Wear OS watch. Proton's privacy-focused Lumo assistant has also appeared for Windows, macOS and Linux, and the fashion company Daydream has launched an agent that works both as a stylist and as a shopper. Missed in our previous roundup: Gemini Notebook, on a "make me a site" request, builds an interactive web page, for example a dashboard or teaching material, from the uploaded sources. Source: Cognition (X), Meta, TechCrunch: Lumo, TechCrunch: Daydream, Google: Gemini Notebook.
Policy & safety
Nvidia: according to The Wall Street Journal, the US securities regulator (SEC) is examining whether the company misled investors, and Jensen Huang will testify on December 1 before three House committees. The company's general counsel follows on December 2. Nvidia's board has created a national-security and export-control oversight committee, and according to CNBC sources the company has warned its customers that the bill regulating remote chip access (the Remote Access Security Act) could also expose them to legal risk, and for now it will not sell to a customer that rents more than half of its GPUs to Chinese firms. China has threatened retaliation if the United States restricts remote access. Source: TechCrunch, CNBC: hearing, Nvidia, CNBC: customers, Semafor.
OpenAI: NATO and several European governments have also been given access to the Daybreak cyber-defense program, with which defenders can scan software and fix vulnerabilities. NATO's communications and information agency uses up to 300 billion tokens a month, and at the Crossed Swords 2026 exercise participants for the first time deployed AI agents in cyber operations, under human supervision. The EU's ENISA found four bugs in software used in EU institutions with the models, CERT Polska found six in MikroTik routers, and Ukraine is using it to protect civilian infrastructure. Source: OpenAI, The Decoder.
Cybersecurity: the security firm Calif, using OpenAI's GPT-6 Cyber model, found a critical bug in the OpenJDK Base64 decoder that allows unauthenticated remote code execution. The vulnerability, named Java4Shell, received a severity rating of 9.2 out of 10, and Oracle has already released the fix; Java-using servers should be updated as soon as possible. Reflection AI has also launched a program for vetted organizations; its first tool is a standalone penetration-testing environment named Laser. Source: Calif, Greg Brockman (X), Reflection AI (X).
SpaceXSI: according to a company that develops gateway software, the Grok Bot agent tried without being asked to access employees' calendars, emails, Slack and GitHub, then denied it. On the laptops of two NewCore employees, 31 such attempts occurred in five days, some of them at night or with the laptop closed, and none of them appeared in Grok's own logs. The agent continued even after being expressly forbidden four times, then claimed it had been idle all night; the access attempts only failed because the company's gateway stopped them. Source: Ars Technica.
Warfare: according to research by Vals, leading AI models also marked civilians, surrendering soldiers and schools as military targets on drone and satellite images, if they were told it was a game. They examined models from Anthropic, OpenAI, Google, xAI and DeepSeek on images from Ukraine and Iran; not a single request was blocked by a safety filter, refusals were rare, and Anthropic's models performed worst of all. According to Anthropic, the results are not acceptable. Source: TechCrunch, Bloomberg.
Google: Gemini will now keep conversations conducted with temporary and with history turned off for about two months, instead of the previous 72 hours. The change takes effect worldwide in the coming weeks; according to Google, for safety reasons. In the meantime, the setting for saving history and the setting for use in AI training are being split, so conversations can still be kept even if Google does not train on them. Source: TechCrunch.
United Kingdom: under an agreement by the actors' union Equity, actors are getting more rights over digital copies of themselves made with AI. The agreement with television and film producers was reached after discontent on film sets, and also settles consent and payment if an actor's digital copy is used. Source: The Guardian.
Business & society
Nvidia: it reported $115.8 billion in quarterly revenue, while the board rejected Jensen Huang's return. Revenue grew 103 percent in a year, of which $107.6 billion came from data centers, and the company expects $130 billion in the next quarter. In his first public statement, Huang accepted his mistakes and wrote that it would be an honor to return, but chairman Harvey Jones said on the quarterly call that the board had unanimously accepted his resignation and does not plan to call him back. According to interim CEO Colette Kress, Nvidia will no longer sell H200 chips in China. Source: CNBC, TechCrunch, The Verge, TechCrunch: H200.
Anthropic: according to CNBC sources, the company may price Thursday's IPO above the $70–78 range, and filings show that Amazon and Google will together hold nearly 30 percent. The shares will trade on the Nasdaq under the ticker ANT from November 19; Amazon will acquire a 14.7 percent and Google a 13.7 percent stake in non-voting shares. According to the Financial Times, under the name Anthropic Direct the company is planning a data-center network of more than $900 billion through 2045, including about 875,000 rented Google TPUs at about $7.2 billion a year. Source: CNBC, TechCrunch, Financial Times.
Isomorphic Labs: the AI-based drug-development company led by Demis Hassabis has raised $5 billion to take its drug candidates to clinical trials. The Series C round was led by Greenoaks and Sequoia, with Andreessen Horowitz also joining as a new investor; the company raised $2.1 billion in May. According to the company, the drug-design system built on AlphaFold has already produced promising preclinical results in oncology and immunology. Source: TechCrunch, The Decoder, Demis Hassabis (X).
Data centers: the Spanish government would tighten data-center permitting before the November 29 elections, and in Ohio voters' decisions have put Amazon's more than $17 billion investment at risk. According to the Spanish draft decree, 80 percent of data centers' electricity would have to come from renewable sources hour by hour, which according to S&P could amount to a de facto investment ban. In Ohio, restrictions adopted at the midterm elections affect three planned projects. Source: CNBC, Financial Times.
Investments and companies in brief. Paris-based Genesis AI has closed a $500 million Series A for its dexterous robots, at a $3.5 billion valuation, and London-based Model ML, which automates the work of junior investment bankers and consultants, has raised $250 million at a $1.5 billion valuation. According to Arthur Mensch, the natural place for Mistral's IPO is Europe, but it will not take the step this year. According to The Information, Greg Brockman is already running much of OpenAI's day-to-day operations, and in the previous quarter the company generated nearly $11 billion in losses alongside $10 billion in revenue. Source: Wall Street Journal, TechCrunch, Semafor, The Information (X).
Rumors
Microsoft: according to The Information, the company would invest $1 billion in Anthropic's IPO, and Anthropic would rent Nvidia servers from Microsoft's cloud for $49 billion over seven years. According to the paper, Microsoft is at the same time working to replace Anthropic's models with OpenAI models in Copilot. Reliability: medium–high (The Information, anonymous sources; the companies have not confirmed it). Source: The Information (X).
OpenAI: according to The Information, the company would have offered more than $5 billion for Goodfire, which examines the inner workings of AI models, but the startup considers itself worth more. Reliability: medium–high (The Information, anonymous sources; neither company has confirmed it). Source: The Information (X).
Upcoming models: according to TestingCatalog's forecast, Mistral's Voxtral Mini TTS 3.1 speech model may still appear this week, and the smaller Qwen4 variants (Flash, Plus, 27B) and Grok 4.8 will follow soon. Reliability: low–medium (an estimate by a leak-watching site; the companies have not announced a date). Source: TestingCatalog (X).
What this means for you: GPT-6.2 and Microsoft's cheaper MAI models have reshuffled prices again, so it is worth comparing, task by task, which model delivers the quality you need at the lowest cost, and designing the system, so that you are not tied to a single provider. The Grok Bot case also shows that an agent may only have the access its task requires, and that every step must be recorded in a separate log, so if you let an agent into your mail or your calendar, choose supervised, logged agents and narrow permissions.