AI News December week 1

Cling AI's "Nano Banana" & Native Audio Cling AI made five days of announcements, including Cling 01 (described as the "Nano Banana of video" for its ability to edit/combine video, images, and text) and Cling Video 2.6, which introduces native audio generation. Testing showed mixed results; while some visual transfers worked (e.g., tablet replacing a book), the physics and specific audio prompts often missed the mark . Runway Gen 4.5 Tease Runway teased Gen 4.5, a new video model that appears to outperform competitors like Sora and VO3.1 in visual quality. Although not yet available for public or early testing, demo videos show impressive consistency and complex scenes . Google Workspace Studio (AI Agent Builder) Google introduced Workspace Studio, a tool similar to Zapier but for building AI agents within the Google ecosystem. It allows users to create workflows with triggers (e.g., receiving an email) and actions (e.g., labeling, asking Gemini to summarize, and notifying via chat). Gemini 3 Deepthink Availability Google's most capable model, Gemini 3 Deepthink, is now available to users on the Gemini Advanced Ultra plan ($250/month). It reportedly crushes benchmarks on tasks like "Humanity's Last Exam" and Arc AGI 2, though it may sometimes be unavailable due to high demand. Amazon Reinvent Announcements (Agents & Chips) At its Reinvent event, Amazon unveiled Frontier Agents for software development and security (e.g., Hero, DevOps agents). They also announced the Tranium 3 chip, optimized for cost-efficient training, and the AI Factory, which brings AI infrastructure (like Nvidia GPUs) to customers' on-premise data centers . DeepSeek v3.2 & v3.2 Special DeepSeek released two new "reasoning-first" models: v3.2 and v3.2 Special. These models achieve state-of-the-art results comparable to major competitors but are trained much more cheaply and efficiently. The "Special" version is highlighted for its agentic capabilities. Mistral 3 Family Open Sourcing French lab Mistral released the Mistral 3 family (3B, 8B, and 14B parameters) under the Apache 2.0 license. These models perform on par with DeepSeek and other top models but distinguish themselves by being fully open-source for local use and fine-tuning . OpenAI "Garlic" Rumors OpenAI has a new model codenamed "Garlic", rumored to be a version of GPT-5 (e.g., GPT-5.2). It is reportedly performing exceptionally well in coding and reasoning tasks and might be released as early as early next year . Ads Coming to ChatGPT Code references and beta user reports indicate OpenAI is testing ads within ChatGPT. Some users on paid plans have already seen search-style ads (e.g., for Target) appearing below responses, sparking concerns about the user experience . Apple's Starflow V Apple demonstrated Starflow V, a new video generation model that uses a normalizing flow-based method rather than the standard diffusion method. This approach offers benefits like faster training and native multitask support (text-to-video, image-to-video, video-to-video) .