Welcome to the "AI Daily" section! This is your guide to exploring the world of artificial intelligence every day. Each day, we present you with the latest content in the AI field, focusing on developers to help you understand technical trends and innovative AI product applications.
Fresh AI products click to learn more:https://app.aibase.com/zh
1. DeepSeek releases V3.2-exp model, introducing a groundbreaking sparse attention mechanism that cuts API costs by half
DeepSeek released a new experimental model called V3.2-exp, significantly reducing the cost of long-context operations through an innovative 'sparse attention' mechanism. The model combines the 'lightning indexer' and 'fine-grained token selection system', improving the efficiency of processing long context segments and showing a 50% reduction in API call costs in preliminary tests.
AiBase Summary:
⚡ DeepSeek launches the V3.2-exp model, using a sparse attention mechanism to optimize long-context processing.
🔍 The lightning indexer and fine-grained token selection system work together to enhance model efficiency.
💰 Preliminary tests show a 50% reduction in API call costs, providing a more economical solution for AI applications.
2. Anthropic makes a big move! Claude Sonnet4.5 outperforms GPT-5, creating a new king in coding
Anthropic released the Claude Sonnet4.5 model, which excels in coding tasks and complex task handling, becoming one of the best coding models available. Its performance improvements are significant, supporting multi-platform use, and enhancing security and alignment.
AiBase Summary:
✅ Claude Sonnet4.5 performs well in coding benchmark tests, achieving over 30 hours of autonomous working time.
🔧 New features such as checkpoints, context editing, and memory tools improve development efficiency and practicality.
🔒 Emphasizes security, reduces risky behaviors, and is suitable for high-risk enterprise scenarios.
3. ChatGPT: Chat and Buy! AI Revolution in E-commerce: One-click Ordering, No Need to Switch Browsers
ChatGPT introduced the 'Instant Checkout' feature, allowing users to complete single-item purchases directly within the chat interface without redirecting to links or browsers. This feature is powered by the 'Agent Commerce Protocol' developed by OpenAI and Stripe, supports multiple payment methods, and will expand to multi-item shopping carts and international markets.
AiBase Summary:
💡 ChatGPT introduces the 'Instant Checkout' feature, enabling direct ordering within the chat interface.
🔒 The 'Agent Commerce Protocol' ensures secure, simple transactions compatible with multiple payment methods.
🌐 Future expansion to multi-item shopping carts and international markets will enhance user experience.
4. OpenAI to Launch an AI Version of TikTok: All Content Created by AI Across the Web
OpenAI is set to launch a social application based on the Sora2 model, called the 'AI Version of TikTok', where all content is generated by AI. The app is designed similarly to TikTok but limits video durations to 10 seconds and supports user identity authentication and portrait usage. Additionally, OpenAI focuses on safety and copyright issues to enhance user experience and prevent user attrition.
AiBase Summary:
🎥 Sora2 generates videos limited to 10 seconds, focusing on concise content dissemination.
🔒 Users can authenticate their identities, and Sora2 can use their portraits for video generation, allowing other users to tag them.
🛡️ OpenAI will send reminders to ensure users are aware when their portraits are used, while addressing copyright issues.
5. Claude Code 2.0 Surpasses Expectations: Checkpoints + VS Code Plugin, Programming Efficiency Increases Threefold
Anthropic released updates to Claude Code v2.0 and Claude Sonnet4.5 models, significantly enhancing AI's autonomy and integration in programming. Claude Code improves the development experience with checkpoint mechanisms, terminal and IDE optimization, and API extensions for developers.
AiBase Summary:
✅ Claude Code v2.0 introduces checkpoint functionality, allowing AI to automatically save states and support rollbacks, enhancing development security.
🔧 The VS Code native extension is entering beta testing, offering inline difference previews and graphical interactions to improve collaboration efficiency.
📈 Sonnet4.5 scored 61.4 on the OSWorld benchmark test, performing exceptionally well, especially in building complex agent systems.
6. Baidu Map Upgrades Xiaodu Think 2.0: Smart Travel Assistant Evolves Completely
Baidu Map launched Xiaodu Think 2.0 at the 7th World New Energy Vehicle Conference, this version being the first industry-end-to-end voice language large model deeply integrated, providing users with more intelligent and personalized travel services. Its core advantages include: introducing a map travel knowledge base and real-time search data to improve understanding of complex travel intentions; building cross-end memory to achieve seamless connection across multiple devices; and having immediate, recent, and long-term memory capabilities to provide personalized recommendation services.
AiBase Summary:
🚗 Baidu Map introduces a map travel knowledge base and real-time search data, improving the ability to understand and reason about complex travel intentions.
📱 Cross-end memory enables seamless connectivity across mobile phones, car systems, and other scenarios, enhancing user experience.
🧠 The smart assistant has immediate, recent, and long-term memory capabilities, providing personalized recommendation services.
7. Ant Group Opens Source the World's First Trillion-Parameter Large Model Ring-1T-preview
Ant Group's Ring-1T-preview is the world's first open-source trillion-parameter inference large model, performing excellently in multiple tests, surpassing several known open-source models, and approaching GPT-5. The model demonstrates strong capabilities in natural language reasoning and code generation, and the team is conducting further training to explore its potential further.
AiBase Summary:
🌟 The world's first open-source trillion-parameter inference large model, Ring-1T-preview, is released.
🚀 Performs well in AIME25 and CodeForces tests, approaching the level of GPT-5.
🧠 The team is conducting post-training to enhance the model's natural language reasoning capabilities.
8. DeepMind Launches the "Frame Chain" Concept: Video Models May Achieve Full Visual Understanding
DeepMind's 'Frame Chain' (CoF) concept brings breakthrough progress to video generation models. This technology enables video models to perform reasoning in both time and space, showcasing general abilities similar to language models. The Veo3 model performs well in multiple visual tasks, demonstrating strong perception, modeling, and control capabilities.
AiBase Summary:
🎥 Frame chain technology allows video models to have cross-temporal and spatial reasoning capabilities, enhancing the intelligence level of video generation.
🧠 Veo3 model demonstrates strong general visual capabilities, capable of handling various untrained tasks.
🚀 DeepMind predicts that future general video models may replace specialized models, driving a new era in machine vision.
Details link: https://papers-pdfs.assets.alphaxiv.org/2509.20328v1.pdf
