This week’s AI news highlights major releases like Claude Opus 4.6 and GPT-5.3 Codex, alongside powerful open-source models such as Step 3.5 Flash and Qwen3 Coder, all pushing the boundaries of coding, reasoning, and multimodal capabilities. Significant advancements in video editing, animation, and robotics—including deepfake tools and real-world robot learning—demonstrate the rapid evolution and expanding accessibility of AI technologies for creators and developers.
This week in AI has seen a surge of groundbreaking releases across both closed and open-source models, as well as new tools for video, coding, and robotics. Major highlights include Anthropic’s release of Claude Opus 4.6, their smartest model yet, which excels in knowledge work, agentic coding, and reasoning, outperforming previous versions and competitors on several benchmarks—though it is slower and more expensive. Almost immediately after, OpenAI launched GPT-5.3 Codex, a powerful coding agent that not only surpasses previous models in autonomous coding tasks but also played a role in its own development through recursive self-improvement. Both models are currently available through paid plans, with API access coming soon for Codex.
On the open-source front, several state-of-the-art models have been released. Step 3.5 Flash is a medium-sized, highly efficient model that rivals much larger proprietary models in reasoning and research tasks, while MiniCPM 0.4.5 stands out as an omnimodal model capable of understanding and responding to text, images, audio, and video in real time. Intern S1 Pro, a trillion-parameter mixture-of-experts model, delivers top-tier performance in scientific research, outperforming even leading closed-source models in specialized domains. Alibaba’s Qwen3 Coder Next is a compact yet powerful coding agent, highly efficient for agentic coding tasks and tool usage, and is already integrated with platforms like OpenClaw.
Video and animation tools have also seen significant advancements. New AI frameworks now allow for transferring the movement from a reference video onto a new character, with the added ability to control camera movement in 3D space. Skin Tokens can estimate the skeletons of 3D models—including non-human characters—for animation, while Omnimat Zero enables object removal and background swapping in videos with impressive accuracy. Edit Yourself is a powerful video editing AI that can seamlessly change what a person says in a video, adjust lip sync, and remove or add content, raising both creative possibilities and concerns about deepfakes.
In robotics, notable progress includes the Uni G1 robot learning to skateboard using the Husky framework, demonstrating real-time adaptation to unstable environments, and enduring extreme cold in a 130,000-step walking challenge. Interact Avatar from Tencent enables AI-generated avatars to interact with objects and perform complex, timed actions based on text prompts, surpassing previous animation tools by allowing real-world object manipulation. InterPrior is another innovation, training robots in virtual environments to interact with objects naturally, which can then be transferred to real-world applications.
Finally, new video generation and editing tools are pushing creative boundaries. Context Forcing allows for the generation of consistent, long videos, while Fast VMT and FS Video offer rapid, high-quality video motion transfer and generation. Google’s Paper Banana automates the creation of accurate academic diagrams and illustrations, outperforming human designers in several aspects. These advancements, along with the release of highly capable OCR models like GLM OCR, highlight the rapid pace of AI development and the growing accessibility of powerful tools for creators, researchers, and developers alike.