The Week in AI: November 17-23, 2025 (Week 47)

News of the week: November 17-23, 2025 (Week 47)

The week was dominated by Google, which released the Gemini 3 model, the Antigravity development environment and the Nano Banana Pro image generator. Anthropic secured commitments of up to $15 billion in investment from Microsoft and Nvidia, and its Claude models arrived on Azure. Nvidia reported record quarterly revenue of $57 billion, yet fears of an AI bubble continued to weigh on tech stocks. On the regulatory front, the European Commission proposed delaying parts of the AI Act, and a draft order emerged in the US aimed at curbing state AI laws.

New Models

Google Releases Gemini 3

On Tuesday, November 18, Google unveiled Gemini 3, which it calls its most capable model yet. Gemini 3 Pro was available from day one in the Gemini app, in AI Mode in Search, and for developers in AI Studio and Vertex AI (Fortune).

According to figures published by Google, the model tops the LMArena leaderboard with a score of 1501 points. On the demanding Humanity’s Last Exam benchmark, it scored 37.5% without tools and 41% in the extended Deep Think mode (Neowin).

Google said Deep Think mode, in which the model “thinks” longer about complex tasks, would come to Google AI Ultra subscribers only after further safety testing. The release came less than a week after OpenAI launched GPT-5.1.

OpenAI: GPT-5.1-Codex-Max for Long Coding Tasks

On Wednesday, November 19, OpenAI released GPT-5.1-Codex-Max, a model built for agentic coding, where the AI works on code independently over extended periods. It replaced GPT-5.1-Codex as the default model in the Codex tool (OpenAI).

The key new feature is a technique called compaction: the model continuously condenses its working history so it can handle tasks that exceed a single context window. OpenAI promised API access at a later date.

xAI Releases Grok 4.1

On Monday, November 17, xAI made Grok 4.1 available to all users on grok.com, on X and in its mobile apps, including the free tier. The company cites fewer hallucinations (made-up answers) and better “emotional intelligence” in conversation (TechRepublic).

Open Models: Olmo 3 and SAM 3

November 20 – The Ai2 institute released Olmo 3, a family of fully open models in 7 and 32 billion parameter sizes under the Apache 2.0 license. It also published the Dolma 3 training data, the code and intermediate training checkpoints (GeekWire).
November 19 – Meta introduced SAM 3, a model that finds, outlines and tracks objects in images and video based on a text prompt. It is complemented by SAM 3D, which creates a 3D model of an object or person from a single photo (SiliconANGLE).

Products and Agents

Google Antigravity: Coding with Agents in the Lead Role

Alongside Gemini 3, Google launched the Antigravity development environment on November 18. Unlike conventional editors with an AI assistant, its main interface is an AI agent to which the developer hands off entire tasks and then monitors their progress (InfoWorld).

Users choose how much autonomy to give the agent. In addition to Gemini 3 Pro, the environment also supports Anthropic’s Claude Sonnet 4.5 and the open GPT-OSS model.

Nano Banana Pro and Image Verification with SynthID

On Thursday, November 20, Google released the Nano Banana Pro image generator (officially Gemini 3 Pro Image). Compared to its predecessor, it renders more legible text in images, supports resolutions up to 4K and handles reference images better (Android Central).

The same day, Google added a feature to the Gemini app that checks whether an image was created or edited by Google’s AI. The app looks for the invisible SynthID watermark, but it cannot reliably detect images from other tools (Google Blog).

Microsoft Ignite: Agent 365 for Managing AI Agents

At the Microsoft Ignite conference (starting November 18 in San Francisco), Microsoft unveiled Agent 365. It serves as a control center where companies register their AI agents, manage their access rights and monitor security using Defender, Entra and Purview (Microsoft Book of News).

More App News

November 20 – OpenAI rolled out group chats in ChatGPT worldwide. Up to 20 people can be invited to a conversation, with GPT-5.1 Auto responding; users’ personal memory is not used in groups (Engadget).
November 19 – ChatGPT for Teachers: verified K-12 teachers in the US get a dedicated version of ChatGPT free until June 2027. Data from it is not used for training by default (OpenAI).
November 17 – Alibaba launched a public beta of its Qwen app, built on its own models. According to the company, more than 10 million people downloaded it in the first week (Alibaba Group).

Business and Investment

Microsoft and Nvidia Invest in Anthropic

On Tuesday, November 18, Microsoft, Nvidia and Anthropic announced a strategic partnership. Nvidia is set to invest up to $10 billion and Microsoft up to $5 billion (Bloomberg).

Anthropic committed to purchasing $30 billion of compute capacity on Azure and to using up to one gigawatt of capacity on Nvidia hardware. This made Claude the only frontier model available on all three major clouds, though AWS remains the company’s primary partner (Axios).

Analysts point to the “circular” nature of such deals: the investor also wins a large contract from the company it is investing in.

Nvidia: Record Revenue, Nervous Markets

On Wednesday, November 19, Nvidia reported results for the third quarter of fiscal 2026. Revenue reached $57 billion, up 62% year over year, of which $51.2 billion came from data centers. The company expects roughly $65 billion for the next quarter (Nvidia).

Even so, the shares fell after an initial rise, and fears of an AI bubble returned (Business Today). Markets had been sliding since the start of the week (Fortune).

Adding to the mood, Google CEO Sundar Pichai told the BBC on November 18 that he sees “elements of irrationality” in the market. If the bubble bursts, he said, no company will be immune, including Google (Malay Mail).

Jeff Bezos Returns to an Executive Role

According to a New York Times report on November 17, Jeff Bezos has become co-CEO of the startup Project Prometheus, which has raised $6.2 billion. It is his first executive role since stepping down as head of Amazon in 2021 (The Register).

The other co-CEO is Vik Bajaj, formerly of Google X and Verily. The company aims to apply AI to engineering, manufacturing and scientific research.

More Deals and Infrastructure

November 19 – The US Department of Commerce approved the export of up to 35,000 Nvidia Blackwell chips each to G42 (UAE) and Humain (Saudi Arabia), subject to strict security conditions (The National).
November 19 – Brookfield launched a $100 billion AI infrastructure investment program with Nvidia and Kuwait’s sovereign wealth fund KIA. At its core is a fund targeting $10 billion in equity (BNN Bloomberg).
November 20 – OpenAI and Foxconn agreed to jointly design server racks and other data center hardware manufactured in the US. OpenAI gets the option to buy the systems, with no purchase commitment (Bloomberg).
November 18 – Intuit signed a multi-year deal with OpenAI worth more than $100 million. The TurboTax, QuickBooks and Credit Karma apps will appear directly in ChatGPT (TechCrunch).

People

November 19 – Yann LeCun, Meta’s chief AI scientist and a Turing Award winner, confirmed he will leave at the end of the year after 12 years. He will found a startup focused on Advanced Machine Intelligence, meaning systems that understand the physical world and can plan (heise online).
November 19 – Economist Larry Summers resigned from the OpenAI board after Congress released his email correspondence with Jeffrey Epstein (CBS News).

Regulation and Law

EU Proposes Delaying Strict AI Act Rules

On Wednesday, November 19, the European Commission presented the Digital Omnibus package, intended to simplify European digital rules. It includes a proposal to delay the AI Act rules for high-risk systems from August 2026 to December 2027 (Biometric Update).

These cover, for example, AI used in hiring, creditworthiness assessment, healthcare or biometric identification. The package also touches on GDPR and cookie rules; the proposal still needs approval from member states and the European Parliament.

The White House Prepares to Take On State AI Laws

According to a draft order dated November 19, reported by Reuters and other outlets, President Donald Trump was considering creating an AI Litigation Task Force at the Department of Justice. It would challenge individual states’ AI laws in court (FedScoop).

States with “overly burdensome” rules could see some federal funding restricted. Critics argue that only Congress, not an executive order, can establish federal preemption.

Security and Incidents

Cloudflare Outage Takes Down ChatGPT and X

On Tuesday, November 18, Cloudflare, the network that protects and speeds up a large part of the web, suffered a major outage. For several hours, services including ChatGPT, X and Canva were down (PC Gamer).

The cause was not an attack but a faulty change to a database query in the Bot Management system. It generated an oversized configuration file that crashed the core proxy servers (GIGAZINE). The incident once again showed how heavily AI services depend on a handful of infrastructure providers.