【AI日报】2026-06-25

追踪

────────────────────

[01]Cursor AI

- It's a demonstration of Cursor's IDE showing AI-powered coding assistance features. The interface is

- Auto 模式:允许Cursor 自动选择模型通常是不扣除信用额度的。 二、 Cursor vs 其它主流 AI 工具 在2026 年,开发者不应只盯着一个工具,而是要了解它们的“特长”。     

[02]Claude Fable 5

- Anthropic发布Claude Fable 5和Claude Mythos 5两款大模型。Fable 5配备完整安全过滤层，Mythos 5移除网络安全领域的安全限制，仅供审查通过的网络防御者和关键基础设施运营商使用。

- Claude Fable 5发布后遭美国官方封禁。该模型因其全球顶级性能和能力引发政策关注，触发了监管部门的干预措施。

[03]AI编程工具

- Claude Code在2026年依然保持长上下文处理优势，在处理跨几十个文件的复杂需求时逻辑一致性领先，很少出现前后逻辑对不上的问题，相比新出工具更适合大型项目重构。

- 小米MiMo Code正式开源，基于OpenCode构建的AI编程智能体，通过SQLite FTS5持久化项目记忆库解决传统工具上下文遗忘问题，支持长线自动化和复杂持续开发任务。

[04]Claude AI

- Anthropic推出Claude Tag，一款面向企业团队的AI协作助理。用户可在Slack中@Claude调用，让其读取上下文、调用授权工具并处理团队协作任务，首先在Slack和Salesforce平台推出。

- Claude Tag是标准Claude Code的演进版本，更主动且更适合团队协作。团队可为其配置特定频道的访问权限，并将其连接到所选工具和信息源。

[05]AI Agent 2026

- 2026年6月，AI Agent商店生态全面爆发。GPT Store、扣子、Dify等主流平台同步推出Agent商店，零代码开发者已实现规模变现，部分开发者月收入超4.7万元，电商客服Agent等应用售价299元已售出300+份。

- AI Agent技能系统成为核心能力建设方案。Skills技能系统为Agent提供动态的专业能力扩展机制，替代传统硬编码方式，提高了Agent的灵活性和可复用性，成为构建高级Agent的关键技术。

模型发布

────────────────────

[01]OpenAI and Broadcom unveil LLM-optimized inference chip

- OpenAI与Broadcom推出Jalapeño芯片，优化LLM推理性能和能效。

[02]Introducing computer use in Gemini 3.5 Flash

- Google DeepMind在Gemini 3.5 Flash中推出计算机交互功能。

工具

────────────────────

[01]Accelerating Transformers Fine-Tuning with NVIDIA NeMo AutoModel

https://huggingface.co/blog/nvidia/accelerating-fine-tuning-nvidia-nemo-automodel

- NVIDIA NeMo AutoModel加速Transformer模型微调效率。

[02]Introducing the FFASR Leaderboard: Benchmarking ASR in the Real World

https://huggingface.co/blog/ffasr-leaderboard

- FFASR排行榜发布，提供真实场景语音识别性能基准测试。

研究

────────────────────

[01]RIFT-Bench: Dynamic Red-teaming For Agentic AI Systems

- RIFT-Bench框架用于智能体系统的动态红队测试与安全评估。

[02]Neuro-Symbolic Drive: Rule-Grounded Faithful Reasoning for Driving VLAs

- 神经符号驱动模型提升自动驾驶VLA的可解释推理能力。

[03]Critique of Agent Model

- 深入分析智能体的定义和特性，审视LLM智能体系统的本质。

[04]Safe and Generalizable Hierarchical Multi-Agent RL via Constraint Manifold Control

- 约束流形控制方法实现多智能体安全协作学习。

[05]Reinforcement Learning Towards Broadly and Persistently Beneficial Models

- 强化学习跨域泛化对齐技术提升模型长期有益性。

[06]Systematic Exploration of 4-Expert Heterogeneous Mixture-of-Experts via Automated Pipeline Search

- 自动化搜索异构四专家混合模型架构优化方案。

行业动态

────────────────────

[01]Europe's extreme heat is shutting down power plants

https://www.technologyreview.com/2026/06/24/1139676/europe-heat-power-plants/

- 欧洲创纪录高温导致电力需求飙升，部分发电厂被迫关停。

[02]The Download: introducing the Engineering issue

https://www.technologyreview.com/2026/06/24/1139658/the-download-introducing-engineering-issue/

- MIT Tech Review推出工程专题，探讨技术改造世界的雄心。

[03]Stripe, Anthropic, and OpenAI are backing an effort to stop respiratory infections

https://www.technologyreview.com/2026/06/24/1139621/stripe-anthropic-and-openai-are-backing-an-effort-to-stop-respiratory-infections/

- Stripe等科技企业联手资助呼吸道感染防治研究。

[04]The emergence of the web data infrastructure layer for AI

https://www.technologyreview.com/2026/06/24/1139202/the-emergence-of-the-web-data-infrastructure-layer-for-ai/

- 网页数据基础设施层崭露头角，解决AI模型训练数据获取瓶颈。

[05]This flying solar-powered platform could deliver better internet from the air

https://www.technologyreview.com/2026/06/24/1138771/solar-powered-platform-delivers-better-internet/

- Sceye太阳能飞行平台可从高空提供互联网覆盖。

[06]Mistral launches OCR 4, turning document extraction into a full enterprise AI play

https://venturebeat.com/data/mistral-launches-ocr-4-turning-document-extraction-into-a-full-enterprise-ai-play

- Mistral OCR 4升级文档智能，提供结构化提取和置信度评分。

[07]Alibaba's model never trained as an agent — and improved agent performance across seven benchmarks

https://venturebeat.com/technology/alibabas-model-never-trained-as-an-agent-and-improved-agent-performance-across-seven-benchmarks

- 阿里Qwen-AgentWorld无需智能体训练即优化七域智能体性能。

[08]Xiaomi's HarnessX rewrites its own AI scaffolding mid-task — and smaller models gain the most

https://venturebeat.com/orchestration/xiaomis-harnessx-rewrites-its-own-ai-scaffolding-mid-task-and-smaller-models-gain-the-most

- 小米HarnessX动态优化智能体框架，小模型性能提升明显。

[09]Stanford researchers will discuss their agentic 'scientists' that are on course to reshape drug discovery at VB Transform 2026

https://venturebeat.com/data/stanford-researchers-will-discuss-their-agentic-scientists-that-are-on-course-to-reshape-drug-discovery-at-vb-transform-2026

- 斯坦福智能体模型有望变革药物发现流程，降低90%失败率。

[10]How Shopify built an AI stack that doesn't care which models survive

https://venturebeat.com/orchestration/how-shopify-built-an-ai-stack-that-doesnt-care-which-models-survive

- Shopify构建LLM代理支持多模型切换，提供自动故障转移。

[11]Cerebras stock plunges after earnings as CEO says margin outlook was misunderstood

https://techcrunch.com/2026/06/24/cerebras-stock-plunges-after-earnings-as-ceo-says-margin-outlook-was-misunderstood/

- Cerebras上市后首份财报利润预期不达预期导致股价下跌。

[12]AI was supposed to kill engineering jobs, but new data suggests they're the most resilient

https://techcrunch.com/2026/06/24/ai-was-supposed-to-kill-engineering-jobs-but-new-data-suggests-theyre-the-most-resilient/

- 数据显示工程岗位反而成AI时代最具韧性的职位。

[13]AI researchers continue to leave Google for its rivals

https://techcrunch.com/2026/06/24/ai-researchers-continue-to-leave-google-for-its-rivals/

- 多位顶级AI研究人员离开谷歌加入Anthropic。

[14]The memory chip crunch is paying off for this US company

https://techcrunch.com/2026/06/24/the-memory-chip-crunch-is-paying-off-for-this-u-s-company/

- 美企芯片业务受益于产业紧缺，营收和利润同比大幅增长。

[15]Companies are scrambling to stop employees from maxing out AI budgets with small tasks

https://techcrunch.com/2026/06/24/companies-are-scrambling-to-stop-employees-from-maxing-out-ai-budgets-with-small-tasks/

- 企业部署Token配额限制，应对员工滥用AI服务成本问题。

[16]How to Opt Out of Google Search's New AI Data Training Feature

https://www.wired.com/story/how-to-opt-out-of-google-search-new-ai-data-training/

- 谷歌搜索新增媒体上传数据用于AI训练，用户可选择退出。

[17]A24 Knows You're Mad About the Google AI Collab

https://www.wired.com/story/a24-knows-youre-mad-about-the-google-ai-collab/

- 谷歌DeepMind投资A24电影公司引发电影爱好者反感。

[18]I Met With China's Top AI Experts. They're Freaking Out, Too

https://www.wired.com/story/ai-arms-race-china-us-cooperation/

- 中美AI竞赛加剧，两国研究者均担忧科技失控风险。

[19]The Trump White House Is Over Anthropic CEO Dario Amodei

https://www.wired.com/story/the-trump-white-house-is-over-anthropics-dario-amodei/

- Anthropic首席执行官在白宫高管会议中被同创人替换。

[20]Qualcomm Buys Buzzy Chip Startup Modular for Nearly $4 Billion

https://www.wired.com/story/qualcomm-buys-buzzy-chip-startup-modular-for-nearly-dollar4-billion/

- 高通近40亿美元收购AI芯片软件新秀Modular。

付费速览

────────────────────

[01]Claude Code (参考)

- 计费模式：按用量计费（Usage-based）

- 方案说明：通过 Anthropic API 计费，无独立订阅，按 token 消耗付费

- 价格详情：Claude Sonnet 4: $3/M input, $15/M output；Claude Opus 4: $15/M input, $75/M output

- 官网：查看最新定价

[02]Cursor (参考)

- 计费模式：订阅制 + 用量计费

- 方案说明：Hobby 免费（2000次/月补全）；Pro $20/月（无限补全 + 500次高级请求）；Business $40/用户/月

- 价格详情：高级模型（GPT-4o、Claude）按请求计费，超出套餐后按 token 收费

- 官网：查看最新定价

[03]Trae (参考)

- 计费模式：免费 + 订阅制

- 方案说明：Free 版每月有限额度；Pro 版按月订阅，含更多模型调用次数

- 价格详情：字节跳动出品，国内版免费额度较高，海外版定价参考官网

- 官网：查看最新定价

[04]Codex (参考)

- 计费模式：ChatGPT 订阅内含 + API 按 token 计费

- 方案说明：包含在 ChatGPT Plus($20/月)、Pro($200/月)、Teams($25/用户/月) 中；也可 pay-as-you-go

- 价格详情：codex-mini-latest API：$1.50/M input，$6/M output；Plus 用户有任务次数限制

- 官网：查看最新定价

[05]Qoder (参考)

- 计费模式：订阅制（Credits 体系）

- 方案说明：Pro / Pro+ / Ultra 三档；折扣已于 2026-04-30 结束，恢复原价

- 价格详情：Teams：$40/用户/月（3000 Credits/月）；个人版三档按月订阅，具体价格见官网

- 官网：查看最新定价