Key week in the global race for artificial intelligence, marked by three major fronts: the competition between Anthropic 4.6 and GPT-5.3 Codex, the offensive of Chinese models like GLM-5 and MiniMax 2.5, and the unstoppable acceleration of research. However, the true protagonist has been Seedance 2.0, ByteDance's new text-to-video model, which promises to revolutionize the audiovisual industry by generating videos of up to two minutes in 1080p, with perfectly synchronized images and sound (dialogues, music, effects) and narrative coherence between scenes and characters. The multilingual lip-syncing and realism quality place Seedance 2.0 among the best systems in the world, reducing costs and simplifying traditionally complex processes. In the realm of open models, the Chinese startup z.ai has introduced GLM-5, an open-source model under MIT license, aimed at businesses and with one of the lowest hallucination rates recorded. GLM-5 stands out for its ability to refrain from responding when unsure, making it especially reliable in professional environments. Additionally, it can generate documents in formats like Word, PDF, or Excel and execute long tasks thanks to its integrated agent functions. Its size (over 700 billion parameters) requires powerful infrastructure, so most will access it via the cloud, where it competes in price with rivals like Claude. Alongside MiniMax 2.5, GLM-5 strengthens the competitiveness of Chinese models against Western ones, both in power and cost and practical applications. OpenAI has begun displaying ads in ChatGPT for free users and those on the Go plan in the U.S., clearly labeled and tailored to the query topic. Users can hide or remove them by subscribing to higher plans. This move brings ChatGPT closer to the search engine model and opens the debate on the impact of advertising on trust and usage of the tool. In the legal arena, Adam Mosseri, Instagram's director, stated in a U.S. trial that social media is not "clinically addictive," although he acknowledges it can create dependency similar to a TV series. The case, a pioneer in the country, seeks to demonstrate that the design of Instagram and YouTube fosters addiction and affects mental health, while Meta and YouTube argue that the plaintiff's issues stem from family origins and that YouTube is not a social network. The trial has revealed internal Meta decisions regarding beauty filters and potential risks for teenagers. Elon Musk has announced a major restructuring of xAI following the departure of two co-founders, dividing the company into four areas: Grok (chatbot and voice), Coding, Imagine (video generation), and Macrohard (AI software with digital agents). The merger with SpaceX, valued at $1.25 trillion, aims to strengthen funding for data centers and talent. xAI is expanding its supercomputer in Memphis with an investment of over $20 billion, betting on real-time video generation and the development of new applications like X Chat and X Money. Musk has reiterated that Grok will not have ads. Anthropic, for its part, plans to secure up to 10 gigawatts of capacity in data centers, with investments of hundreds of billions of dollars and hiring former Google executives to gain control over hardware and reduce dependence on external suppliers. Access to energy and computing is consolidating as the new battleground of AI. Alibaba has introduced RynnBrain, a model that equips robots with vision and coordination to interact with the physical environment, expanding the Qwen family into robotics and betting on open source to accelerate adoption. The competition in physical AI is intense, with Nvidia, Google DeepMind, and Tesla developing similar systems. OpenAI has launched GPT-5.3 Codex, specialized in software development and automation, which stands out for its autonomy and efficiency in complex tasks, and has introduced an official app to coordinate projects with AI agents, bringing development closer to a more orchestrated and less manual model. Codex Spark, its ultra-fast version, reinforces the trend towards more autonomous and efficient programming, while Anthropic bets on deploying multiple agents, albeit with higher resource consumption. In research, Anthropic has presented an "interpretability microscope" to analyze how models like Claude process information, revealing the existence of a "universal language of thought" and strategies for anticipatory planning. Meanwhile, a joint study by MIT, Meta FAIR, and NYU proposes SOAR, a meta-learning scheme where a "teacher" generates problems and a "student" solves them, demonstrating that the quality of the problems is key to unlocking learning in models with low training signal. OpenAI and Ginkgo have connected GPT-5 to an autonomous lab, allowing the model to propose experiments, execute them, and learn from the results, achieving a 40% reduction in protein production costs after six iterations and over 36,000 compositions tested. Other news: Gemini 3 Deep Think achieved 84.6% on ARC-AGI-2 and gold medals in science olympiads; ChatGPT is preparing to introduce skills; subscriptions to X exceed $1 billion annually; new Siri features are delayed until iOS 26.5 or 27; the device from Jony Ive and OpenAI is postponed to 2027; AirPods 4 will have cameras and cost $249; Musk prioritizes the Moon over Mars; MiniMax 2.5 matches OpenAI and Anthropic in agent systems at a much lower cost; and Aurora's autonomous trucks now surpass humans in distance traveled.