The Week in AI: December 8-14, 2025 (Week 50)

News of the week: December 8-14, 2025 (Week 50)

The week was dominated by OpenAI’s response to pressure from Google: the company released its GPT-5.2 model and landed a billion-dollar investment from Disney. Washington cleared exports of Nvidia H200 chips to China, and President Trump signed an order meant to curb AI regulation at the state level. Rivals in AI agents joined forces under a single foundation for the first time, with Anthropic handing over its MCP protocol. Meanwhile, the stock market showed just how nervous investors are about massive data center spending.

New Models

OpenAI Releases GPT-5.2

On December 11, OpenAI unveiled GPT-5.2 in three variants: the fast Instant, the reasoning-focused Thinking, and the most powerful Pro. The model went first to paying ChatGPT users, with free users getting access gradually (InfoWorld).

According to OpenAI, the Thinking version matched or beat human experts in 70.9% of tasks on the GDPval benchmark (which compares AI work with professionals across 44 occupations). That said, the benchmark was created by OpenAI itself. In the API, the model costs $1.75 per million input tokens and $14 per million output tokens, more than the previous GPT-5.1 (VentureBeat).

Media outlets link the release to an internal “code red” that leadership declared over the success of Google‘s Gemini 3. OpenAI executives, however, deny that it was merely a reaction to the competition.

Mistral Targets Developers with Devstral 2

On December 9, France’s Mistral AI released Devstral 2, a family of coding models in two sizes: 123 billion parameters and the smaller Devstral Small 2 with 24 billion. Put simply, parameters are the model’s “adjustable numbers”; the more there are, the more capable the model usually is, but also the more demanding on hardware (Mistral AI).

Alongside the model came Mistral Vibe CLI, an open source coding agent controlled from the command line. With it, Mistral competes directly with tools such as Anthropic’s Claude Code.

December 8 – Chinese company Zhipu AI (Z.ai) released the open models GLM-4.6V and the smaller GLM-4.6V-Flash, which understand images and can call tools on their own. They are available under the MIT license (VentureBeat).
December 11 – Runway introduced its first “world model,” GWM-1, which simulates environments frame by frame, physics included. Its Gen-4.5 video model can now also generate audio and dialogue (TechCrunch).
December 12 – Google upgraded its Gemini 2.5 Flash Native Audio voice model, which now follows instructions better and calls functions more reliably. It is rolling out in Gemini Live and to developers (The Decoder).

Products and Agents

Rivals Launch a Foundation for AI Agents

On December 9, the new Agentic AI Foundation was established under the Linux Foundation. Its founders are Anthropic, OpenAI, and Block, with support also coming from Google, Microsoft, AWS, Cloudflare, and Bloomberg (Anthropic).

Anthropic donated the MCP (Model Context Protocol) to the foundation; it has become the standard for connecting AI to external tools and data. The other founding projects are Block’s goose agent and OpenAI’s AGENTS.md format, which tells agents how to work with a specific codebase.

Google Opens Its Deep Research Agent to Developers

On December 11, Google launched an improved Gemini Deep Research agent that searches the web on its own and compiles detailed research reports. It runs on Gemini 3 Pro, and for the first time developers can build it into their own applications (9to5Google).

This is done through the new Interactions API, a unified interface for both models and agents, currently in public beta (Google).

Google Translate Interprets Right in Your Headphones

On December 12, Google Translate launched a beta of live speech translation that users hear directly in their headphones. It works with any headphones, supports more than 70 languages, and is powered by Gemini models (TechCrunch).

For now it is available only on Android in the US, Mexico, and India. It is expected to reach iOS and more countries in 2026.

December 8 – At The Android Show: XR Edition, Google showed off smart glasses with Gemini that it is developing with Warby Parker and Gentle Monster. The first units are due in 2026 (Vision Monday).
December 10 – Adobe made its Photoshop, Express, and Acrobat apps available for free in ChatGPT. Users can now edit a photo or PDF right in the chat, except in the EU and the UK (Adobe).
December 9 – The Pentagon launched the GenAI.mil platform with Gemini for Government for roughly 3 million civilian and military personnel and contractors (Fox Business).

Business and Investment

Disney Puts a Billion into OpenAI

On December 11, Disney announced it will invest $1 billion in OpenAI and also receive warrants for additional shares. The deal includes a three-year license that will let users of the Sora video app create videos featuring more than 200 characters from Disney, Marvel, Pixar, and Star Wars (CNN).

The license does not cover the likenesses or voices of real actors. Selected videos are set to appear on Disney+, with the first results expected in early 2026.

Oracle and Broadcom: Record Demand, Nervous Investors

On December 10, Oracle reported quarterly results with revenue of $16.1 billion, slightly below expectations. What spooked investors most was a hike in planned data center spending to roughly $50 billion for the fiscal year (Fortune).

Oracle shares fell about 11% on December 11. The market is worried about how quickly the spending will pay off, and the company is building a large share of its capacity for OpenAI.

The same day, Broadcom reported revenue of $18 billion and 74% year-over-year growth in AI chip sales. CEO Hock Tan revealed that the previously unnamed customer behind a $10 billion order is Anthropic, which has ordered another $11 billion. Even so, the stock dropped after the results due to an expected decline in margins (TrendForce).

December 8 – IBM will acquire Confluent, a specialist in real-time data streaming, for about $11 billion. The “live” data streams are meant to feed enterprise AI systems (Nasdaq).
December 9 – Accenture and Anthropic signed a multi-year partnership. Around 30,000 of the consulting firm’s employees will be trained on Claude, and its developers will get Claude Code (TechCrunch).
December 9 and 10 – Microsoft announced a $17.5 billion investment in AI and cloud in India through 2030, and a day later Amazon pledged $35 billion in the country (The Week).

Chips and Geopolitics

Nvidia Can Sell H200 to China, Washington Wants 25%

On December 8, Donald Trump announced that Nvidia may sell H200 chips to approved customers in China. The US government will take 25% of those sales (The Register).

The newer Blackwell chips and the upcoming Rubin remain banned. A similar arrangement is expected to apply to AMD and Intel; China hawks have criticized the move.

On December 10, Reuters reported that Nvidia has developed software that estimates which country a chip is running in based on communication delays between servers. It is meant to help combat chip smuggling; the company stresses that it is voluntary and contains no “kill switch” (SCMP / Reuters).

Regulation and Law

Trump Moves to Curb State AI Laws

On December 11, President Trump signed the executive order “Ensuring a National Policy Framework for Artificial Intelligence.” It aims to replace the patchwork of individual state laws with a single, minimally burdensome federal framework (Bloomberg Law).

The Attorney General is to set up a task force that will challenge state laws in court. States with “excessive” regulation could lose part of their federal funding for internet expansion. Exceptions are planned for areas such as child protection.

State Attorneys General Warn About “Delusional” Chatbot Responses

On December 10, 42 US state attorneys general sent a letter to thirteen companies, including OpenAI, Google, Microsoft, Meta, and Anthropic. They warn that sycophantic and “delusional” chatbot responses may violate state laws (TechCrunch).

The attorneys general are calling for safety testing before models are released, independent audits, and user notifications similar to those for data breaches. They gave the companies until January 16, 2026 to commit.

EU Investigates How Google Uses Content for AI

On December 9, the European Commission opened an antitrust investigation into Google. It is examining whether the AI Overviews and AI Mode features use publishers’ content and YouTube videos without fair compensation and without a way to opt out (Engadget).

If a violation is proven, Google could face a fine of up to 10% of its global annual turnover. Google argues that the complaint stifles innovation.

Safety and Incidents

OpenAI: Upcoming Models Will Be “High Risk” in Cybersecurity

On December 10, OpenAI warned that its upcoming models are likely to reach the “high” level for cybersecurity under its Preparedness Framework. That is the second-highest tier; the models could, for example, help find previously unknown vulnerabilities (Axios).

The company plans to invest in defensive tools and set up a Frontier Risk Council, a board of outside security experts.

Around the World

December 11 – Time magazine named the “Architects of AI” its 2025 Person of the Year, honoring the people behind the rise of artificial intelligence, among them Jensen Huang, Sam Altman, and Dario Amodei (Global News).