AI News Roundup: December 13 – December 24, 2025

The most important news and trends

December 16, 2025

Google pilots ‘CC’ AI email assistant via Labs

Google launched a new experimental AI assistant called “CC” through Google Labs, an email-based agent that connects with Gmail, Drive, and Calendar. CC provides a daily briefing email summarizing the user’s schedule, tasks, and updates, and can handle commands via email (like adding to-do items or notes). The pilot is limited to North American AI Pro and Ultra plan users (consumer accounts only) and aims to streamline productivity by proactively surfacing relevant info each morning. Why it matters: It shows Google embedding generative AI deeper into personal productivity tools, experimenting with new assistant formats beyond chatbots to drive user engagement in everyday workflows.

Source: TechCrunch

Adobe Firefly adds prompt-based video editing and new AI models

Adobe updated its Firefly generative AI platform with a suite of new video creation features and model integrations. A new beta AI video editor allows users to make precise edits to generated clips using text prompts (leveraging Runway’s Aleph model) and to apply custom camera movements. Firefly also integrated Topaz Labs’ Astra model for video upscaling up to 4K and Black Forest Labs’ FLUX.2 for photorealistic image generation. Additionally, Adobe is temporarily offering unlimited image/video generations for paying Firefly subscribers to encourage use of the new tools. Why it matters: Adobe is bolstering its generative AI toolkit for creators, combining its in-house models with specialized third-party AI to expand capabilities—signaling a collaborative approach to advancing creative AI and keeping Adobe’s ecosystem competitive as AI content tools proliferate.

Source: TechCrunch

OpenAI announces Apple Music integration for ChatGPT

OpenAI revealed that Apple Music will be among the new third-party services integrated into ChatGPT’s app directory. Once live, users will be able to ask ChatGPT to create music playlists or find songs via natural language prompts, similar to the existing Spotify plugin. The announcement, made by OpenAI’s apps chief Fidji Simo, comes as the company opens submissions for ChatGPT Apps and expands its roster of integrations (which already includes Spotify, Expedia, Zillow and more). Why it matters: The move underscores ChatGPT’s evolution into a broader AI platform, as OpenAI courts major partners to extend its chatbot’s utility—transforming ChatGPT from a pure Q&A tool into a hub that can directly interact with popular services and content.

Source: 9to5Mac

ChatGPT mobile apps add ‘branched chat’ feature

OpenAI’s ChatGPT app for iOS and Android introduced a new “Branch in new chat” option, allowing users to split off any message into a separate conversation thread. The branched chats feature, which launched on the web version in September, lets users explore different questions or directions without cluttering a single long conversation. Mobile users can now long-press a message to start a branch, aligning the app with the desktop experience and making complex or multi-topic interactions more manageable. Why it matters: This update improves the usability of ChatGPT on mobile, giving users more control to organize and experiment with AI dialogues—key for productivity and creative workflows—while maintaining context across branches, which was previously only possible on the desktop client.

Source: TechRadar

December 17, 2025

Amazon in talks to invest ~$10 billion in OpenAI

Amazon.com is in discussions to invest about $10 billion into OpenAI, which would value the ChatGPT creator at over $500 billion. The negotiations are fluid, but a deal could involve OpenAI using Amazon’s in-house AI chips (Trainium) and selling a tailored version of ChatGPT for Amazon’s use. This comes after OpenAI’s recent restructuring and a $38 billion cloud contract with Amazon in November, and indicates OpenAI’s willingness to partner beyond its primary backer Microsoft. Why it matters: A partnership of this scale would dramatically deepen Amazon’s involvement in generative AI while providing OpenAI with massive capital and cloud resources — underscoring how tech giants are racing to secure alliances and infrastructure in the AI boom.

Source: Reuters

Elon Musk’s xAI opens Grok Voice Agent API to developers

Musk-founded xAI released the Grok Voice Agent API, enabling outside developers to build voice-based AI agents using xAI’s in-house speech technology. The API exposes the same real-time voice stack that powers Grok in Tesla vehicles and xAI’s apps, supporting dozens of languages and rapid tool-calling for tasks like web searches. xAI touts Grok Voice’s speed (under 1 second to first audio) and cost efficiency ($0.05 per minute) as highly competitive, and has optimized the system for natural-sounding voices and multilingual interactions. Why it matters: This marks xAI’s bid to challenge incumbents in AI voice assistants by leveraging its integration with Tesla and low-cost model — potentially pressuring rivals like OpenAI and Google on real-time AI and signaling Musk’s ambition to expand his AI ecosystem beyond text-based chatbots.

Source: xAI (company blog)

Google launches Gemini 3 Flash, a faster AI model for its apps

Google introduced Gemini 3 Flash, a new lightweight version of its Gemini AI model optimized for speed and cost-efficiency. Despite a smaller footprint, Gemini 3 Flash achieves performance on par with larger “Pro” models on many benchmarks, narrowing the gap between quick replies and deep reasoning. Google made Flash the default model in its consumer Gemini app and search AI, replacing the previous 2.5 Flash, while still allowing users to switch to the more powerful model for complex tasks. Enterprise partners (like Figma, Harvey, JetBrains) are already using Gemini 3 Flash via Google’s cloud services. Why it matters: The launch of Gemini 3 Flash highlights Google’s strategy to offer AI experiences that are both fast and capable, aiming to undercut rivals by reducing latency and cost. Making it the default for millions of users raises the baseline for AI assistants and intensifies competition with OpenAI’s models in consumer and enterprise applications.

Source: TechCrunch

December 18, 2025

OpenAI launches GPT-5.2-Codex, an advanced AI coding model

OpenAI unveiled GPT-5.2-Codex, a specialized version of its GPT-5.2 model tailored for “agentic” software engineering tasks. The model is optimized to handle long coding sessions, large-scale code refactoring, and cybersecurity use cases, outperforming previous Codex iterations on benchmarks for terminal-based tasks and coding reliability. OpenAI rolled out GPT-5.2-Codex to paying ChatGPT users on launch and plans to extend it to API customers, while also piloting enhanced access for vetted cybersecurity professionals given the model’s powerful capabilities. Why it matters: GPT-5.2-Codex represents a leap in AI-assisted programming, indicating how rapidly AI can take on complex, long-horizon coding and security analysis. Its release underscores OpenAI’s push into professional domains and raises dual-use concerns, as increasingly capable code-generation AI could both bolster software development and introduce new security considerations.

Source: OpenAI (company blog)

Google’s Gemini app can verify AI-generated videos

Google rolled out a feature in its Gemini AI app that allows users to check whether a given video was created or edited using Google’s AI. By uploading a video and querying “Was this generated using Google AI?”, the app will look for Google’s SynthID watermarks across audio and visuals and report where AI content is detected (e.g., “SynthID detected in audio between 10–20 seconds”). The tool works on clips up to 90 seconds and is available globally across all languages supported by Gemini, aiming to boost transparency amid growing concerns over deepfakes. Why it matters: As AI-generated media proliferates, Google providing a built-in authenticity checker is a significant step toward combating misinformation. It reflects tech companies’ increasing responsibility to help users discern AI-altered content, using watermarking and detection to uphold trust in digital media.

Source: The Verge

OpenAI opens ChatGPT App Store to third-party developers

OpenAI began allowing developers to submit third-party “ChatGPT apps” for review and listing in a new App Directory inside ChatGPT. The submission portal went live on Dec 17, letting external apps (beyond the initial set of partners) integrate into ChatGPT so that users can discover and activate them within conversations. OpenAI will vet all apps for compliance and safety, and approved apps will roll out to ChatGPT’s 800+ million users in early 2026. This expansion builds on the ChatGPT SDK introduced in October and significantly broadens the chatbot’s plugin ecosystem, with dozens of new apps (Adobe, Gmail, Replit, etc.) already added beyond the original few (e.g., Spotify, Expedia). Why it matters: This marks ChatGPT’s transformation into a full-fledged platform, not just an AI assistant. By opening an “app store” model, OpenAI is fostering a developer ecosystem that can embed specialized tools and services directly into AI dialogues — a move poised to accelerate ChatGPT’s usefulness and monetization, but also one that raises new questions about oversight and data privacy in AI-augmented workflows.

Source: VentureBeat

Perplexity brings Google’s Gemini 3 Flash model to its AI search

AI search startup Perplexity announced that the newly launched Gemini 3 Flash model from Google is now fully available to its Pro and Max subscribers. Gemini 3 Flash is a lightweight, high-speed language model that delivers low-latency responses without sacrificing much accuracy. By integrating this model, Perplexity aims to provide faster, more cost-efficient answers and better handle complex, multi-turn queries in real time, leveraging Flash’s strong language understanding and lower inference costs. Subscribers don’t need to opt in — the system will automatically use Gemini Flash when appropriate, based on the query type. Why it matters: This move illustrates how third-party AI services are quickly adopting state-of-the-art models from big AI labs to stay competitive. Perplexity’s use of Gemini 3 Flash highlights the demand for quicker, cheaper AI inference in consumer applications and underscores Google’s influence in distributing its models across the AI ecosystem beyond its own platforms.

Source: Pandaily

December 19, 2025

Google launches FunctionGemma, a small on-device AI assistant model

Google released a new AI model named FunctionGemma, a 270-million-parameter specialized model that runs on devices to translate natural language commands into executable actions. Unlike cloud-based chatbots, FunctionGemma is designed to operate locally (on smartphones, browsers, IoT devices) as a “router,” instantly handling user requests like app controls or navigation without an internet round-trip. Google open-sourced the model via HuggingFace and Kaggle, and provided developers with a full recipe (model weights, dataset, and tooling support) to adapt it for their own apps. The model emphasizes privacy (data stays on-device), low latency, and no per-call API costs, heralding a shift toward “small language models” for edge use. Why it matters: FunctionGemma reflects Google’s strategic pivot toward more efficient, private AI deployments on consumer devices. By empowering phones and browsers with capable mini-models, Google is challenging the notion that only giant cloud AI models are useful — potentially broadening AI’s reach and setting new expectations for speed, offline functionality, and cost in everyday AI assistants.

Source: VentureBeat

ByteDance cloud unveils Doubao 1.8 AI model as usage soars

ByteDance’s cloud arm, Volcano Engine, announced an upgrade to its flagship AI model, Doubao 1.8, alongside a new multimodal creation model called Seedance 1.5 Pro. The company said Doubao 1.8 has reached “global top-tier” status in multimodal understanding and agent capabilities, doubling the frames it can analyze in video inputs and improving tool-use for complex tasks. President Tan Dai also reported that Doubao’s average daily usage now exceeds 50 trillion tokens – more than ten times last year’s level – with over 100 enterprise clients each accumulating 1+ trillion tokens of usage, making Doubao one of the most-used cloud AI models in China. Why it matters: The scale and advancement of Doubao underscore China’s rapid progress in AI – ByteDance is not only achieving massive model adoption domestically but also pushing technical boundaries in multimodal AI. This highlights intensifying global competition, with Chinese tech firms scaling up AI usage and capabilities to rival Western models in both performance and sheer volume of real-world use.

Source: Pandaily

December 23, 2025

AI data centers keep old ‘peaker’ power plants online

A Reuters investigation finds that surging electricity consumption from AI data centers is delaying the retirement of dozens of aging fossil-fuel “peaker” power plants in the U.S. These plants, intended for occasional peak use, are being run more often to meet data centers’ round-the-clock demand, reversing earlier plans to shut them down. Peaker plants emit more pollution per unit of power than typical plants and tend to be located in low-income and minority communities, raising environmental justice concerns as regulators consider keeping them operational to avoid AI-related grid shortfalls. Why it matters: It reveals an unintended consequence of the AI boom: the race for more computing power is straining electrical grids and undermining clean energy goals. Ensuring stable power for AI expansion may come at the cost of increased local pollution and carbon emissions, putting pressure on policymakers to balance tech growth with environmental and public health priorities.

Source: Reuters

NYT reporter Carreyrou sues OpenAI, Google, xAI over AI training

John Carreyrou, a New York Times investigative reporter and author of “Bad Blood,” filed a lawsuit (with five other authors) accusing OpenAI, Google, Elon Musk’s xAI, Meta, Anthropic, and Perplexity of misusing their copyrighted books to train AI models. The suit, filed in a California court, alleges the companies “pirated” the texts without permission, and notably marks the first legal action to name xAI as a defendant. Unlike prior class-action cases by authors (one of which Anthropic settled for $1.5B), Carreyrou’s group is suing individually, aiming to avoid class settlements and seek up to $150,000 per infringed work in damages. Why it matters: This lawsuit escalates the growing conflict over AI training data and intellectual property. It highlights prominent journalists and authors directly challenging AI companies, potentially setting important legal precedents about whether using copyrighted material to train AI violates the law – a question at the heart of AI’s impact on creative industries.

Source: Reuters

YouTube tests ‘Playables Builder’ to create mini-games from prompts

YouTube launched a closed beta of “Playables Builder,” an AI-powered tool that lets creators generate simple web games using text, image, or video prompts. Powered by Google’s Gemini 3 model, the prototype web app can turn a short game description or reference media into a playable game (without coding), which creators can then share directly on the YouTube platform. YouTube provided examples of games made with the system and is inviting select creators to sign up as trusted testers; the pilot is limited to certain markets and testers will receive separate credentials to access the builder. Why it matters: This experiment signals YouTube’s ambitions to expand beyond video into interactive content using generative AI. By enabling game creation for non-developers, YouTube could unlock a new form of user engagement and content monetization, while also showcasing the versatility of Google’s Gemini models in creative tasks like game design.

Source: 9to5Google