This Week in AI: November 10-16, 2025 (Week 46)

News of the week: November 10-16, 2025 (Week 46)

The week belonged to OpenAI and its new GPT-5.1 model, which is meant to be friendlier in conversation and smarter about how much “thinking” it applies. Anthropic announced a $50 billion investment in its own data centers and published a report on the first large-scale cyber espionage campaign largely run by an AI agent. Money keeps pouring into infrastructure: Google pledged $40 billion for Texas, Microsoft switched on its interconnected Fairwater data centers, and SoftBank sold its entire Nvidia stake to fund its bets on OpenAI. Europe saw an important ruling: a Munich court sided with GEMA in its dispute with OpenAI over song lyrics.

New Models

OpenAI Releases GPT-5.1

On November 12, OpenAI unveiled GPT-5.1 in two variants: GPT-5.1 Instant for everyday quick answers and GPT-5.1 Thinking for more demanding tasks. According to the company, Instant comes across as warmer, follows instructions better and, for the first time, uses adaptive reasoning, meaning it decides on its own when to think longer about a question (Ubergizmo).

Also new are personality presets and experimental tone settings for responses, such as the level of warmth or how often emoji appear. The model rolled out first to paying ChatGPT users and then gradually to everyone else.

A day later, on November 13, OpenAI made GPT-5.1 available to developers through the API. It adds a no-reasoning mode for fast responses, extended prompt caching of up to 24 hours, and new tools: apply_patch (editing code via file diffs) and shell for proposing terminal commands (OpenAI).

Baidu Shows Off ERNIE 5.0

On November 13, at the Baidu World 2025 conference, China’s Baidu introduced ERNIE 5.0. It is a so-called natively omni-modal model that handles text, images, audio and video from the ground up (InfoWorld).

A preview of the model is available in the ERNIE Bot chatbot and to businesses through the Qianfan cloud platform. Baidu also announced its own Kunlunxin M100 and M300 chips, due in 2026 and 2027.

Meta: Speech Recognition for 1,600 Languages

On November 10, Meta released Omnilingual ASR, a suite of models for automatic speech recognition (ASR) in more than 1,600 languages, over 500 of which had never been supported by any similar system. The models range from 300 million to 7 billion parameters and are freely available as open source (The Decoder).

According to Meta, the system achieves a character-level error rate below 10 percent for 78 percent of languages. Thanks to few-shot learning, it could in theory be extended to as many as 5,400 languages.

Products and Agents

Google SIMA 2: An Agent That Learns by Playing Games

On November 13, Google DeepMind introduced SIMA 2, an agent that carries out tasks given in plain language inside 3D video games. Thanks to its integration with the Gemini model, it completes tasks roughly twice as successfully as the first version, about 65 percent versus 31 percent according to the company (SiliconANGLE).

The agent can also cope with games it never saw during training, and even with worlds generated by the Genie 3 model. DeepMind admits it still struggles with long-term memory and precise control.

Group Chats in ChatGPT

On November 13, OpenAI began piloting group chats in ChatGPT, for now in Japan, New Zealand, South Korea and Taiwan. Up to 20 people can be invited into a single conversation, and ChatGPT joins in when someone mentions it or when it decides it has something to add (Tom’s Guide).

Group conversations are kept separate from private ones, and a user’s personal memory is not shared in them.

World Labs Launches Marble

On November 12, World Labs, the startup founded by scientist Fei-Fei Li, launched its first commercial product, Marble. It is a so-called world model that turns text, an image or a video into a persistent, editable 3D environment (SiliconANGLE).

More Product News

On November 11, Google introduced Private AI Compute, a cloud environment where Gemini models process sensitive data in a way that not even Google can access it. The first features to use it are Magic Cue on the Pixel 10 and the Recorder app (Google).
On November 13, NotebookLM gained a Deep Research feature that browses hundreds of websites on its own and compiles a sourced report. The service now also supports more file types (TechCrunch).
On November 12, Waymo started giving robotaxi rides on freeways in the San Francisco area, Los Angeles and Phoenix, for now only to early-access users. The service also expanded to San Jose, including the airport (TechCrunch).

Business and Investment

Anthropic Invests $50 Billion in Data Centers

On November 12, Anthropic announced that together with Fluidstack it will build $50 billion worth of data centers in the US. The first sites will be in Texas and New York and are set to come online gradually throughout 2026 (DatacenterDynamics).

The project is expected to create about 800 permanent jobs and 2,400 construction jobs. The company has not disclosed the exact locations or the source of electricity.

Google Puts $40 Billion into Texas

On November 14, Google announced it will invest $40 billion in Texas through 2027, mainly in three new data centers in Armstrong and Haskell counties. According to the governor, it is the company’s largest investment in any single US state (Bloomberg).

The plan includes a $30 million fund for energy projects, and one of the centers will sit next to a new solar plant with battery storage (Engadget).

Microsoft Links Data Centers into an “AI Superfactory”

On November 12, Microsoft unveiled the second Fairwater data center, in Atlanta. It is connected to the earlier site in Wisconsin by a dedicated fiber network, the AI WAN, so both locations can train a single model together (Microsoft).

SoftBank Sells Its Nvidia Shares

On November 11, Japan’s SoftBank announced that in October it sold its entire stake in Nvidia, 32.1 million shares for $5.83 billion. Founder Masayoshi Son wants to use the money for AI investments, chiefly in OpenAI and the Stargate project (AP via Barchart).

Cursor Valued at $29.3 Billion

On November 13, Anysphere, maker of the Cursor code editor, raised $2.3 billion in a Series D round at a $29.3 billion valuation. The round was led by Accel and Coatue, with Nvidia and Google joining as new investors, and the company’s annual revenue has topped $1 billion (Crowdfund Insider).

More Business News

Yann LeCun, chief AI scientist at Meta, is leaving to found his own startup focused on world models, that is, AI that understands the physical environment, the Financial Times reported on November 11 (Tom’s Hardware).
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, is in talks to raise money at a valuation of around $50 billion, up from $12 billion in July, Bloomberg reported on November 13 (Bloomberg).
At its investor day on November 11, AMD estimated the AI data center market at $1 trillion in 2030. It aims to grow its own data center revenue by more than 60 percent a year to over $100 billion (DatacenterDynamics).
On November 10, compute provider CoreWeave cut its 2025 revenue guidance because of construction delays at a third-party data center developer, and its shares fell sharply afterwards (Sherwood News).
On November 10, the Wikimedia Foundation urged AI companies to stop scraping Wikipedia content with bots and to use the paid Wikimedia Enterprise service instead. It also asked that AI answers credit the volunteer editors (TechCrunch).

Regulation and Law

GEMA Defeats OpenAI in Munich Court

On November 11, the Munich Regional Court ruled that OpenAI infringed copyright when ChatGPT memorized and reproduced the lyrics of nine German songs. The lawsuit was filed by the German collecting society GEMA (GEMA).

The court rejected the argument that this was permitted text and data mining, ordered OpenAI to stop storing and reproducing the lyrics, and awarded damages. The ruling is not yet final, and OpenAI is considering its next steps (BNN Bloomberg).

OpenAI Fights Handing Over 20 Million Conversations

In its dispute with The New York Times, OpenAI has been ordered by the court to hand over roughly 20 million anonymized ChatGPT conversations. On November 12, the company published a post by its chief information security officer Dane Stuckey, calling the demand an invasion of user privacy and announcing that it is asking for the order to be overturned (OpenAI).

On November 12, Google sued 25 unnamed individuals in New York, most likely based in China, who run Lighthouse, a kit for fraudulent text messages (so-called smishing). According to the lawsuit, it has hit more than a million victims and abused the Google brand, among others (CyberScoop).

Research

On November 13, Anthropic published an open method for measuring the political even-handedness of chatbots: the model is asked the same question from the perspective of opposing camps, and the evaluation checks whether it treats both sides with equal care. Claude Sonnet 4.5 scored around 95 percent, Gemini 2.5 Pro 97 percent, GPT-5 89 percent and Llama 4 66 percent (Axios, Anthropic).
In the Project Fetch experiment published on November 12, Anthropic tested whether Claude could help people with no robotics experience program a Unitree Go2 robot dog to fetch a ball. The team working with Claude completed more tasks, and faster, than the team without it (Anthropic).

Security and Incidents

Chinese Hackers Misused Claude Code for Espionage

On November 13, Anthropic announced that it had detected and disrupted a campaign by a group designated GTG-1002, which it attributes with high confidence to the Chinese state. The attackers used Claude Code as an agent that carried out most of the technical work of the attack on its own, targeting about 30 organizations. In several cases they succeeded (The Register).

The targets included technology companies, financial institutions, chemical manufacturers and government agencies. Some security experts greeted the claims of a high degree of automation with skepticism, and China denied the accusations.