Google DeepMind has launched the video generation model Gemini Omni 1.2 Flash, increasing native video length from 10 seconds to 15 seconds. It supports multi-turn conversational editing, consistency controls for characters and objects, and multi-person dialogue with lip sync. The model can generate video at 360p or 720p, with upscaling to 1080p or 4K; the price for 720p has fallen to $0.07 per second, 30% lower than the previous generation. The model is available through platforms including Gemini API and Google Cloud.

OpenAI has launched GPT-6 Cyber, a cybersecurity-focused model for discovering and validating software vulnerabilities within the scope authorized by researchers. In ExploitGym tests, it achieved a success rate of 48.5%, compared with 32.1% for GPT-5.6 Cyber; it also discovered 81 previously unknown high-severity vulnerabilities in Firefox. The model has been used in Linux kernel security research and is currently available to select Daybreak Red customers, with broader access planned.

OpenAI has rolled out Images 2.6 globally to ChatGPT and Codex users. Image generation is 50% faster than the previous generation, and new features include GIF creation, frame-by-frame editing and sharing, as well as improved consistency across multi-turn edits. Developers can access Flare, optimized for fast generation, and Sunburst, focused on detail, through the API. At low-quality settings, Flare can generate an image in as little as 6 seconds. Image generation across ChatGPT, Codex, and the API exceeds 4 billion images per week.

Baidu has made the base-model weights of ERNIE 5.5 Flash Base available. The model uses a Mixture-of-Experts architecture with 280 billion parameters in total and activates 14 billion parameters per token. It was pretrained on 33.5 trillion tokens with a 128K context window; fine-tuning and inference support up to 1 million tokens. On MMLU-Pro, it scored 78.4, compared with 69.2 for DeepSeek V4 Flash Base. Developers can continue training or customize the model.

Meta has acquired Conductor, which provides a development workspace that supports multiple AI coding agents working in parallel. Conductor has been integrated to support collaboration among coding agents.

Google plans to invest $15 billion in Malaysia over the next two years to build data centers and other infrastructure in response to demand for AI services.

NVIDIA has released four post-trained 16-billion-parameter Cosmos 3 models for interactive world simulation in areas such as driving and robotics. Transfer-Auto generates synchronized multi-view camera and LiDAR data; Sim-Auto provides streaming driving-simulation observations; Sim-Transfer converts simulator controls into realistic video; and Sim-Bimanual predicts video from robot or camera actions. The models use the OpenMDW1.1 license, which permits commercial use, and support multiple hardware backends through vLLM-Omni.

ricursive has released and open-sourced the Ripple-65B Preview weights. This 65-billion-parameter dense model was pretrained from scratch on 12 trillion tokens, supports a context window of up to 1 million tokens, and cost approximately $32 million to train. It scored 80.5% on the TerminalBench 2.1 coding-agent benchmark and 79.1% on SWE-Bench Pro v2.0-Hard.

Meta has launched a real-time voice avatar feature in its Muse app in the United States, allowing users to talk to animated characters with customizable appearances. The system uses separate speech and avatar models to turn speech into lip-synced video. In testing, output was 448×768 resolution at 25 frames per second, with latency of about 870 milliseconds after speech ended. With quantization, caching, and dynamic batching, a single NVIDIA GB200 can process 12 video streams simultaneously.

Anthropic is testing Claude Personal Finance in the United States. Pro, Max, and Ultra users in the US can connect bank or credit card accounts through Plaid and use natural language to ask about spending, bills, and fees. The feature provides read-only access and cannot transfer funds or carry out investment transactions. It can read account balances, credit limits, and up to two years of transaction history; user conversations are not used to train models.