This Week in AI: December 1-7, 2025 (Week 49)

News of the week: December 1-7, 2025 (Week 49)

The week was dominated by pressure on OpenAI: Sam Altman declared an internal “code red” so that ChatGPT can hold its ground against the fast-rising Google Gemini 3. Open models got stronger thanks to DeepSeek V3.2 and Mistral 3, and at re:Invent Amazon unveiled its own models, chips and autonomous agents. Anthropic bought Bun and, according to media reports, is preparing for a stock market listing, while Micron is leaving the consumer market because of data centers’ hunger for memory. On the regulatory front, the EU investigation of Meta over WhatsApp and the New York Times lawsuit against Perplexity stood out.

New Models

DeepSeek V3.2 and V3.2-Speciale

China’s DeepSeek released the DeepSeek-V3.2 model and a more powerful variant, V3.2-Speciale, on December 1. Both come with open weights (downloadable model parameters) on Hugging Face, so anyone can run them on their own hardware. DeepSeek API Docs

According to the company, the Speciale version reached gold-medal level at the IMO math olympiad as well as at the ICPC and IOI 2025 programming contests; these figures come from the developer itself. Speciale uses more tokens, does not support tool calling and was only available temporarily via the API until December 15. TechNode

Mistral 3: An Open Family from Small Models to a Large One

France’s Mistral AI introduced the Mistral 3 family under the permissive Apache 2.0 license on December 2. It consists of three smaller Ministral 3 models (3, 8 and 14 billion parameters) designed to run on laptops or phones, plus the flagship Mistral Large 3. Bloomberg

Large 3 uses a mixture-of-experts architecture: out of 675 billion parameters in total, only about 41 billion are active for each query, which saves computing power. It also handles images and a context of 256 thousand tokens.

Gemini 3 Deep Think for Ultra Subscribers

Google made the Gemini 3 Deep Think mode available to Google AI Ultra subscribers in the Gemini app on December 4. It is a “thinking” mode that takes more time on a complex task and tries several approaches at once. Google Blog

According to the published results, it scored 41% on Humanity’s Last Exam without tools and 45.1% on ARC-AGI-2 with code execution. Google delayed the release because of additional safety testing. 9to5Google

Video: Runway Gen-4.5 and Kling O1

December 1: Runway released the Gen-4.5 video model, which took first place on the Video Arena leaderboard by Artificial Analysis, ahead of models from Google and OpenAI. The company highlights more realistic physics and human motion; at launch it only supports text-to-video. eWeek
December 1: China’s Kuaishou launched Kling O1, which combines video generation, editing and retouching in a single model. The company’s shares in Hong Kong rose by more than 5% during the day. eWeek

Products and Agents

AWS re:Invent: Nova 2 Models, the Trainium3 Chip and “Frontier Agents”

At its re:Invent conference in Las Vegas (November 30 to December 4), Amazon Web Services introduced the Nova 2 model family in Lite, Pro, Sonic (voice conversation) and Omni variants. With the Nova Forge service, companies can build their own model on top of Amazon Nova. eWeek

Trainium3 UltraServers are now generally available, featuring Amazon’s own AI training chip, which is meant to be an alternative to Nvidia GPUs. Another new offering is AI Factories, AWS AI infrastructure installed directly in customers’ own data centers. ITPro

AWS also showed three so-called frontier agents: the autonomous programmer Kiro, the AWS Security Agent and the AWS DevOps Agent. They are designed to work on their own for hours or even days without human intervention.

Google Workspace Studio

On December 3, Google launched Workspace Studio, a tool for building AI agents without programming. The user describes a task in plain language and Gemini 3 turns it into an agent that works across Gmail, Drive and Chat as well as with tools such as Jira or Salesforce. Google Workspace Updates, 9to5Google

OpenAI Declares a “Code Red”

OpenAI CEO Sam Altman declared a “code red”, according to an internal memo from December 1 reported by The Information and The Wall Street Journal. The company is to focus on the speed, reliability and personalization of ChatGPT, as competitors led by Gemini 3 are closing the gap. Bloomberg

As a result, work on advertising, shopping agents and the Pulse feature is reportedly being postponed.

Business and Investment

Anthropic Buys Bun as Claude Code Hits a Billion in Revenue

On December 2, Anthropic announced the acquisition of Bun, a fast JavaScript runtime that already powers its coding tool Claude Code. Bun will remain open source under the MIT license. Anthropic

The company also said that Claude Code had reached an annualized revenue run rate of 1 billion dollars just six months after its public launch.

Anthropic Is Preparing for an IPO, According to the FT

According to the Financial Times (December 2), Anthropic has hired the law firm Wilson Sonsini to prepare a possible initial public offering (IPO) as early as 2026. The company said no final decision has been made; at the same time, it is reportedly negotiating a private funding round at a valuation of more than 300 billion dollars. Sherwood News

Micron Abandons the Crucial Brand

Memory maker Micron announced on December 3 that by February 2026 it will shut down its consumer business under the Crucial brand (RAM and SSDs for regular customers). It will shift capacity to large buyers, mainly AI data centers, where memory is in short supply. Sherwood News

Other Deals of the Week

December 1: Nvidia bought 2 billion dollars’ worth of shares in chip design software maker Synopsys (about 2.6% of the company) and expanded their collaboration on simulations and digital twins. Bloomberg
December 1: Accenture made OpenAI one of its key partners and will roll out ChatGPT Enterprise to tens of thousands of employees. OpenAI, Reuters / Yahoo Finance
December 1: Germany’s Black Forest Labs, the maker of the FLUX image models, raised 300 million dollars at a valuation of 3.25 billion dollars. Tech.eu
December 3: OpenAI is buying the Polish startup Neptune, whose tools track the progress of model training; it will shut down the services for external customers. OpenAI
December 3: Snowflake and Anthropic signed a multi-year deal worth 200 million dollars; Claude models will power enterprise data agents for more than 12,600 customers. TechCrunch
December 3: According to The Information, Microsoft lowered its growth targets for AI agent sales; the company denied cutting its overall sales quotas. Its shares temporarily fell by almost 3%. Reuters / Yahoo Finance
December 4: According to Bloomberg, Meta is considering cuts of up to 30% to its metaverse budget for 2026, shifting the money to AI glasses and wearables. Fox Business

Personnel Changes at Apple and Meta

December 1: Apple‘s head of AI John Giannandrea is stepping down and will serve as an adviser until spring 2026. The new vice president of AI is Amar Subramanya, formerly at Microsoft and before that head of engineering for Gemini at Google. Business Wire (Apple)
December 3: Meta poached Apple’s head of interface design Alan Dye, the creator of the Liquid Glass look. At Meta he will lead a new design studio for hardware, software and AI. Bloomberg

Regulation and Law

EU Investigates Meta over AI in WhatsApp

On December 4, the European Commission opened antitrust proceedings against Meta. New WhatsApp Business API rules bar third-party chatbots from the platform if AI is their main service, while Meta’s own Meta AI stays. TechRadar

For existing providers, the rules are set to apply from January 15, 2026. The Commission may order interim measures; WhatsApp called the accusations “baseless”. Reuters / CTV News

The New York Times Sues Perplexity

On December 5, The New York Times filed a lawsuit against Perplexity in federal court in New York. It claims the startup copied millions of articles without permission and attributed made-up answers (hallucinations) to the paper under its trademark. A day earlier, the Chicago Tribune had filed a similar lawsuit. Axios

Ban on State AI Laws Dropped from the Defense Bill

House Majority Leader Steve Scalise acknowledged on December 2 that a provision barring US states from regulating AI on their own would not make it into the NDAA defense bill. Hundreds of state lawmakers had opposed it as well; according to Scalise, Republicans are looking for another way. PYMNTS, Benton Institute

Research

OpenAI Teaches Models to “Confess”

On December 3, OpenAI published research on so-called confessions. The model attaches a second output to its answer in which it reports whether it broke instructions, cheated or guessed, and this output is graded solely on honesty. OpenAI

On a GPT-5 Thinking variant, the method significantly improved the detection of undesirable behavior, but it does not prevent such behavior by itself. According to MIT Technology Review, independent researchers warn that confessions are more of an estimate than a faithful record of the model’s reasoning. MIT Technology Review

Anthropic Interviewer

Anthropic introduced Anthropic Interviewer, a tool in which Claude itself conducts interviews with people about how they use AI. The first study included 1,250 professionals, among them scientists and artists; the company published the interview transcripts as a dataset. Dataconomy (December 5)

Security and Incidents

AI Agents Can Attack Smart Contracts

A study by Anthropic from December 1 tested AI agents on the SCONE-bench benchmark with 405 actually exploited smart contracts (programs on a blockchain that manage money). For contracts exploited after March 2025, the agents “stole” about 4.6 million dollars in simulation, with Claude Opus 4.5, Claude Sonnet 4.5 and GPT-5 performing best. The Block

In 2,849 new contracts with no known bugs, the agents found two previously unknown vulnerabilities (zero-day). It was purely a simulation; Anthropic urges defenders to deploy AI for protection as well. Decrypt