# Have you heard these exciting AI news? - December 12, 2025 AI Updates Weekly

Source: https://www.youtube.com/watch?v=omzgBEhOAHQ
Recap page: https://rapidrecap.app/video/omzgBEhOAHQ
Generated: 2025-12-12T21:05:21.271+00:00

---
## Quick Overview

The December 12, 2025 AI Updates weekly recap covers major news including OpenAI's GPT-5.2 release with tiered pricing, Mistral AI's new Devstral models, the founding of the Agentic AI Foundation supported by Linux, Anthropic, and OpenAI, Google's ongoing integration of Gemini AI across Workspace, deep dives into the Titans Transformer-2 architecture and GLM-4.6 performance, and the announcement of the NeurIPS 2025 conference details.

**Key Points:**
- OpenAI released GPT-5.2 with three tiers (Instant, Thinking, Pro), featuring high input/output tokens and a low hallucination rate, while also developing the 'Garlic' model for early 2026.
- Mistral AI released Devstral 2 (123B params) and Devstral Small 2 (24B params), both open-source and showing strong performance, especially in coding, with Devstral 2 being 7x more cost-efficient than Claude Sonnet 4.5 on coding tasks.
- The Linux Foundation announced the formation of the Agentic AI Foundation, anchored by Anthropic and OpenAI donations, promoting open-source licensing (GNU, Apache, MIT).
- Google is integrating Gemini AI across its Workspace suite (Gmail, Docs, Slides) and launched Google Workspace Studio (formerly Google Workspace Flows) for no-code workflow automation via AI agents.
- DeepSeek v3.2 was released, showing performance close to Claude Sonnet 4 but costing significantly less ($0.60/$2.20 per million tokens vs $3/$15 for Claude Sonnet), and it is open-source.
- The Pentagon ordered the creation of an AI futures steering committee by April 2026 to monitor AI threats and ensure human oversight capabilities over superintelligent systems.
- The presentation highlighted major AI announcements from NeurIPS 2025, including the GPT-5.2 release and the ongoing competition between major AI labs.

![Screenshot at 00:03: The main slide summarizing the key AI updates for December 12, 2025, categorized into sections like LM Arena, HRM, OpenAI/DeepSeek, and US AI Regulations.](https://ss.rapidrecap.app/screens/omzgBEhOAHQ/00-00-03.png)

**Context:** This video provides a weekly summary of significant developments in the AI landscape as of December 12, 2025. Key areas covered include new model releases from OpenAI (GPT-5.2) and Mistral (Devstral), regulatory developments from the US government, advancements in AI integration within Google Workspace, updates on open-source models like GLM-4.6, and the formation of a new open-source AI governance body, the Agentic AI Foundation.

## Detailed Analysis

The AI updates for December 12, 2025, began with OpenAI's release of GPT-5.2, which comes in three model variants (Instant, Thinking, Pro) supporting 400K input and 128K output tokens, boasting a low 1% hallucination rate for Thinking mode and offering significant performance improvements over predecessors. GPT-5.2 pricing is detailed, showing that the Pro version costs $21.00 per million input tokens and $168.00 per million output tokens, which is competitive compared to Claude models. OpenAI is also developing a new model named 'Garlic' for early 2026, expected to outperform Google's Gemini 3 and Anthropic's Claude Opus 4.5. Mistral AI released Devstral 2 (123B parameters, open-source) and Devstral Small 2 (24B parameters), which performed exceptionally well on coding benchmarks, with Devstral 2 being 7x more cost-efficient than Claude Sonnet 4.5. The talk then covered the formation of the Agentic AI Foundation, anchored by the Linux Foundation, Anthropic, and OpenAI, aimed at promoting open standards like MCP and AGENTS.md. In the Chinese AI landscape, several players like Qwen, DeepSeek, Zhipu, and others are actively contributing to open models. Shifting to hardware, China is tripling its AI accelerator output to 500,000 units by 2026, despite US export restrictions, with Michael Dell's 'Invest America' initiative providing $1,000 investment accounts to children born after January 1, 2025. Google continues to integrate Gemini AI deeply into Workspace via the Gemini sidebar and the new Google Workspace Studio (formerly Google Workspace Flows), enabling no-code automation. The update also highlighted DeepSeek v3.2, an open-source model that rivals Claude Sonnet 4, is significantly cheaper on input/output costs, and utilizes an architecture that simplifies maintenance. Finally, regarding regulation, President Trump signed an executive order establishing federal-level AI regulation to prevent conflicting state rules, arguing that 50 states independently regulating AI would harm US competitiveness; the framework excludes child safety, local infrastructure, and copyright issues, which remain under state jurisdiction.

### OpenAI GPT-5.2

- GPT-5.2 released on December 11 with three models (Instant, Thinking, Pro)
- 400K input/128K output tokens
- 1% hallucination rate for Thinking mode
- GPT-5.2 Pro costs $21.00/$168.00 per million tokens
- 'Garlic' model announced for early 2026, outperforming Gemini 3 and Claude Opus 4.5

### Mistral Devstral-2

- Devstral 2 (123B params) and Small 2 (24B params) released, both open-source
- Devstral 2 is 7x more cost-efficient than Claude Sonnet 4.5 on coding
- Outperformed DeepSeek v3.2 in human evaluations with a 42.8% win rate vs 28.6% loss rate

### Agentic AI Foundation

- Formed by Linux Foundation anchoring Anthropic and OpenAI donations
- Focus on open source licensing (GNU, Apache, MIT)
- Includes Chinese AI players like Qwen, DeepSeek, Zhipu, and OpenBMB

### AI in Google Workspace

- Gemini AI available across Gmail, Docs, Slides, etc.
- New Google Workspace Studio (formerly Google Workspace Flows) enables no-code automation via AI agents
- Google aims to build an everything platform while Perplexity focuses on answer quality

### OpenAI's 2025 State of Enterprise AI report

- Significant growth with 8x message volume growth since Nov 2024
- Enterprise AI adoption shifting from isolated usage to structured, workflow-driven integration
- 25% of enterprises still not connected to core data systems

### GLM-4.6

- Open-source model from Zhipu AI released Sept 2025
- Performance close to Claude Sonnet 4
- Costs $0.60/$2.20 per million tokens (7-8x cheaper on output than Sonnet)
- 355B parameters, half the size of DeepSeek-V3

### US AI Regulations

- Trump signed an executive order to prevent conflicting state rules
- Advocates for federal oversight to avoid harming American competitiveness
- Framework excludes child safety, local infrastructure, and copyright issues from federal preemption

![Screenshot at 00:03: Slide header 'AI Updates - Dec 12, 2025' listing key topics for the week.](https://ss.rapidrecap.app/screens/omzgBEhOAHQ/00-00-03.png)
![Screenshot at 01:05: Detailed comparison of crowd-sourced LM Arena leaderboard for English text and Coding tasks, showing model rankings.](https://ss.rapidrecap.app/screens/omzgBEhOAHQ/00-01-05.png)
![Screenshot at 02:31: Slide detailing OpenAI GPT-5.2 features, variants, pricing, and the upcoming 'Garlic' model.](https://ss.rapidrecap.app/screens/omzgBEhOAHQ/00-02-31.png)
![Screenshot at 05:09: Slide detailing Mistral's Devstral-2 models, emphasizing their open-source nature and coding performance benchmarks against competitors.](https://ss.rapidrecap.app/screens/omzgBEhOAHQ/00-05-09.png)
![Screenshot at 07:45: Slide summarizing NeurIPS 2025 conference details, including dates, location, and key research areas.](https://ss.rapidrecap.app/screens/omzgBEhOAHQ/00-07-45.png)