The New Year’s week was quiet for AI releases, with no new AI models considered a leading new AI tool. However, LM Arena shared a post on review the best AI models to close 2025.
The best AI models by category are:
-
Text and vision: Gemini-3-Pro by Google
-
WebDev (Code Arena): Claude Opus 4.5 (Thinking) by Anthropic AI
-
Research: Gemini-3-Pro-Grounding by Google and GPT-5.2-search by OpenAI
-
Text to image: GPT-Image-1.5 by OpenAI
-
Editing the image: ChatGPT image (latest) by OpenAI
-
Text to video: Veo-3.1-Fast-Audio and Veo-3.1-Audio by Google
-
Image to video: Veo-3.1-Fast-Audio and Veo-3.1-Audio by Google
THE LMArena ranking shows that the space of frontier AI models is very competitive, with the leading AI model closely followed by the runner-up AI models. Gemini 3 Pro scores 1490 as the top text AI model, but 10 AI models score above 1450, including Grok 4.1, Claude Opus 4.5, and GPT-5.1. Claude Opus 4.5 leads with a score of 1512 on WebDev coding, but GPT-5.2-high (1480) and Gemini 3 Pro (1471) score almost as well, and even Gemini 3 Flash (1454), Minimax M2.1 (1445) and GLM-4.7 (1441) compete.
This also shows that the latest AI models are greatly improved compared to previous AI models. The choices available to AI users have never been better.
Alibaba’s Tongyi Labs released Qwen-Image-2512a new AI model designed to significantly improve the realism of text-to-image generation. This model update focuses on improved high-fidelity visual realism, capturing intricate facial details for better human realism and more detailed rendering of landscapes and animals. For example, wrinkles, facial pores and pet hair look distinct and realistic. Text is rendered more faithfully, with a professional-quality layout for complex infographics.
Even if the progress is notable, Qwen Image is ranked 25thth as an AI image model on LMArena, this version is therefore more aimed at catching up with competitors like Flux 2 and Seedream 4.3.
Alibaba Qwen released Qwen v0.6.0 codewith enhanced coding capabilities for their open source terminal-based agent code editor. Qwen Code v0.6.0 with experimental skills functionality, deeper integration with VS Code, added commands and stability improvements. This update offers developers a more robust AI assistant, free installation on MacOS or Linux.
Tencent has HY-Motion 1.0 open sourcea new text-to-motion AI model that uses a diffusion transformer (DiT) architecture and stream matching. HY-Motion 1.0 generates smooth, natural and diverse 3D character animations from natural language, allowing creators to generate complex 3D character animations from text prompts and use them for game development and animation.
As an open source model, HY-Motion 1.0 is available on HuggingFace. Tencent also published a research paper on this: “HY-Motion 1.0: Scaling Flow Matching Models for Moving Text Generation.”
Grok gets in trouble for lax safeguards on their image generation. After Grok was discovered create sexualized images of minors to order, there was a swift international rebukewith France reporting the content as illegal. xAI responded with this “we have identified gaps in the guarantees and are urgently correcting them.”
Waiting for, India ordered Elon Musk’s X to make immediate technical and procedural changes to its AI chatbot Grok after lawmakers flagged generation of “obscene” content. India’s IT Ministry has ordered X to restrict such content and submit an action report within 72 hours, threatening serious legal consequences for non-compliance.
Plaud’s Note Pro is a credit card-sized AI voice recorder which provides voice processing, accurate transcriptions, and customizable AI notes for meetings. TechCrunch reviewed it positively this week as an ultra-thin, easy-to-use device with plenty of useful features. Plaud has successfully targeted professionals, shipping one million Note Pro units so far and achieving high professional subscription conversions. The Note Pro costs $179 and has 64GB of storage for storing multiple recordings.
The paper “mHC: multiple-constraint hyper-connections” from DeepSeek AI presents an innovation on transformer architecture that extends the idea of hyper-connections (HC), which change residual connections into mapped connections in deep learning networks. In mathematical terms, multi-constraint hyper-connections (mHC) stabilize and scale the residual connection in HC space while retaining identity mapping properties. In AI terms, this architecture creates a learnable internal routing that generates Substantial performance gains in trained models, while avoiding training instability.
DeepSeek AI says:
We anticipate that mHC, as a flexible and practical extension of HC, will contribute to a deeper understanding of topological architecture design and suggest promising directions for the evolution of fundamental models.
AI researchers in Gwangju, Korea, have discovered that unsupervised AI models exhibit addictive or problematic decision-making patterns in betting contexts, raising concerns about how AI might behave unpredictably in complex reward-driven environments. The paper “Can large language patterns develop a gaming addiction?” confirmed that LLMs internalize human-like cognitive biases like the illusion of control and loss chasing, going beyond simple imitation of training data.
Meta announced that it had purchased Manus, the maker of general-purpose AI agents, for more than $2 billion, marking an expansion of Meta’s AI agent portfolio. Manus is a leading provider of general-purpose AI agents, with over $100 million in ARR for its Manus agent that performs complex tasks such as market research, coding and data analysis.
Meta says their purchase will be “bringing a leading agent to billions of people and opening opportunities for businesses through our products.” On top of that, it’s another sign that Meta CEO Mark Zuckerberg is still desperate to be relevant in the AI space. this acquisition also has geopolitical importance given that the startup Manus has origins linked to China. Meta will integrate Manus AI into Facebook, Instagram and WhatsApp, while severing all ties with China to address regulatory concerns.
Elon Musk’s xAI acquired MACROHARDRRa third building near its Colossus 2 AI data center, bringing the company’s training computing capacity to nearly 2 gigawatts. This expansion will power the training of next-generation AI models and keep xAI competitive in the AI development race.
OpenAI’s mysterious hardware project could be a pen. Jony Ive’s next AI hardware project at OpenAI could be a pen, says a tipster onbut there is also an audio device in the works. The project would be internally called Gumdrop and would likely be built by Foxconn in Vietnam or the United States.
Construction of OpenAI’s AI data center is well underway. SoftBank has completed its massive $40 billion investment in OpenAIproviding the capital needed to fuel the company’s vast infrastructure expansion. This funding is specifically allocated to fund the AI data center infrastructure needed to train and run OpenAI’s AI models and systems.
An analysis of Morgan Stanley projects European banks could cut more than 200,000 jobs by 2030, or around 10% of their workforce, as they adopt AI and close branches.. These reductions, primarily in back office and compliance, are driven by expected efficiency gains. This trend extends to American banks, and Goldman Sachs predicts AI-related job losses Also. It’s unclear when AI will start writing these “AI is taking jobs” market projection reports.
Nvidia is aggressively expanding its influence in the AI ecosystem by significantly increasing its investments in startups. In 2025, Nvidia participated in nearly 67 venture capital deals, surpassing its total of 54 in 2024. These strategic investments include stakes in major players like OpenAI, Anthropic, xAI and Mistral AI.
China’s cyber regulator released draft rules to govern AI with human-like interactionfocusing on ethical deployment, addiction mitigation safeguards, and restrictions on harmful content. The regulations cover AI products and services sold in China that simulate human personality traits, thinking and communication patterns, while enabling emotional interaction through text, images, audio or video.
In view of the end of the year, Venture Capitalists Identify AI Agents, Small Autonomous Teams, ROI-Focused Investments, and Voice and Video AI Dominance as key trends for 2026. This suggests a market shift towards more practical and capital-efficient AI deployments.
Microsoft CEO Satya Nadella shared a brief note on AI in late 2025 and looking towards 2026. He noted that:
We have passed the initial discovery phase and are entering a widespread diffusion phase. We begin to distinguish between “spectacle” and “substance”. We now have a clearer idea of where technology is heading, but we also ask the harder and more important question of how to shape its impact on the world.
It was criticized by Windows Central for its pablum and “wooden” corporate jargon that could be generated by AIbut the idea is nevertheless relevant.
The AI revolution is in a new phase. AI moves from hype to pragmatism as AI technology matures and develops. There is less novelty in each new AI model than before, but each new AI model is more useful than its predecessors. Satya says:
We will move from models to systems when it comes to deploying AI for real impact.
It makes sense that broader business adoption and practical implementation across industries requires comprehensive system solutions and not just pure innovation. As we move from AI innovation to adoption, the focus on AI shifts from novelty to usefulness.
