大模型行业军备竞赛从一级市场的泡沫化高额估值，落地到更底层的算力基础设施建设。 本周又是Agent（智能体）热度持续的一周。北京时间6月19日，国内大模型初创企业Minimax发布能够完成长程（Long Horizon）复杂任务的通用智能体 MiniMax Agent，多步规划专家级解决方案、拆解任务需求、执行多个子任务从而交付最终结果。 大模型干活更强了 在山姆·奥尔特曼的表述中，新模式的显著特性已不是性能多么好，而是干活多么强。 即将发布的GPT-5将不再是一个单一功能的产品，而是一个经过架构融合的综合性产物，该变化折射的是山姆对AGI（通用人工智能）定义理解的变动，他 表示，现在AI的发展已远远超过五年前行业根据软件认知能力而提出的AGI定义。当下更需要考虑的是，AGI需要什么？山姆认为是能够自主发现新科学或 极大提高人们使用工具发现新科学的能力，例如当下正火的AI编程产品。 一位二级市场券商分析师对记者表示，大模型真正产生价值始于Agent开始为人类工作，虽然行业还处于非常早期的阶段。过去的AI也具备提效功能，但存 在功能单一且依赖人类指导、缺乏通用性和自主性等问题。Agent将自然语言转化为具体指令 ...

Core Insights - The price of large models is decreasing due to advancements in AI technology, with OpenAI reducing the price of its o3 model by 80% and Volcano Engine offering significant cost reductions for its video generation model Seedance 1.0 pro [3][4] - OpenAI's price reduction is attributed to comprehensive optimizations in its inference service architecture, and the company is exploring partnerships with Google Cloud to alleviate computing power pressures [3] - Volcano Engine's pricing strategy focuses on the most commonly used input range of 0-32K tokens, with significant cost reductions compared to previous models [4][5] Group 1 - OpenAI's o3 model price cut is a strategic move to enhance competitiveness in the AI market [3] - Volcano Engine's new pricing for its models is based on engineering optimizations and aims to lower inference costs through its AI cloud-native service, ServingKit [4] - The rapid growth in token consumption, particularly in AI search and programming, indicates a strong demand for AI tools and models [5] Group 2 - ByteDance's AI programming product Trae has surpassed 1 million monthly active users, showcasing the practical application of AI coding tools [7] - The evolution of AI agents is expected to transform software from passive tools to active executors, with a focus on deep reasoning and multimodal understanding [7] - The development of protocols like MCP and A2A is crucial for building an efficient agent ecosystem, with Volcano Engine working on next-generation protocols to enhance model tool utilization [8]