日期：2026-04-30
本期聚焦：重点关注模型发布与 release notes、官方 engineering blog、AI coding / agent / SRE、评测榜单变化、开发者实践博客、框架生态、开源模型与真实用户视角；当 HN、Reddit、Hugging Face 等社区源可访问时优先纳入。
-
Artificial Analysis 最新模型排名观察（Artificial Analysis）
中文摘要：Artificial Analysis 最新模型排名显示，GPT-5.5 (xhigh) 以 60 分的智能指数位居榜首，GPT-5.5 (high) 以 59 分紧随其后。Claude Opus 4.7 (Max Effort) 与 Gemini 3.1 Pro Preview、GPT-5.4 (xhigh) 并列第三，均获得 57 分。开源模型方面，Kimi K2.6 以 54 分领跑，MiMo-V2.5-Pro 同分并列，DeepSeek V4 Pro (Reasoning, Max Effort) 以 52 分位列第三。速度方面，Mercury 2 以 778.1 tokens/秒 居首；成本方面，Qwen3.5 0.8B 以每百万 tokens 0.02 美元成为最经济选择。平台目前共评估 367 个模型，其中 232 个为开源权重模型。
English Summary: Artificial Analysis' latest model rankings show GPT-5.5 (xhigh) leading with an Intelligence Index score of 60, followed by GPT-5.5 (high) at 59. Claude Opus 4.7 (Max Effort) ties with Gemini 3.1 Pro Preview and GPT-5.4 (xhigh) at 57 points. Among open weights models, Kimi K2.6 leads with 54 points, tied with MiMo-V2.5-Pro, while DeepSeek V4 Pro (Reasoning, Max Effort) ranks third at 52. Mercury 2 is fastest at 778.1 tokens/s, and Qwen3.5 0.8B is most affordable at $0.02 per 1M tokens. The platform has evaluated 367 models total, including 232 open weights models.
-
Introducing Claude Opus 4.7（Anthropic News）
中文摘要：Anthropic 正式发布 Claude Opus 4.7，该模型在多项企业级基准测试中表现显著提升。在 Rakuten-SWE-Bench 上，Opus 4.7 解决生产任务的数量是 4.6 版本的三倍，代码质量与测试质量均有两位数提升。视觉理解能力大幅增强，在 XBOW 的视觉敏锐度基准测试中从 54.5% 跃升至 98.5%。在 Databricks 的 OfficeQA Pro 测试中，文档推理错误减少 21%。企业用户反馈显示，该版本在智能体决策、工具调用准确性、角色遵循和复杂工程任务协调方面均有明显改善，代码输出也更加简洁，减少了冗余的包装函数。
English Summary: Anthropic officially released Claude Opus 4.7, showing significant improvements across enterprise benchmarks. On Rakuten-SWE-Bench, it resolves 3x more production tasks than Opus 4.6, with double-digit gains in Code Quality and Test Quality. Visual understanding improved dramatically from 54.5% to 98.5% on XBOW's visual-acuity benchmark. On Databricks' OfficeQA Pro, document reasoning errors decreased by 21%.
-
Featured An update on recent Claude Code quality reports（Anthropic Engineering）
中文摘要：Anthropic 工程团队发布 Claude Code 近期质量问题的复盘报告。4 月 16 日随 Opus 4.7 发布时，团队为减少模型冗长输出而添加了系统提示词长度限制（工具调用间文本 ≤25 词，最终回复 ≤100 词），该改动意外导致模型智能下降。此外，一项缓存优化错误地丢弃了历史推理内容，影响代码审查功能，导致模型无法基于先前推理继续工作。团队在收到用户反馈后，于 4 月 7 日将 Opus 4.7 默认 effort 级别恢复为 xhigh，并于 4 月 10 日发布 v2.1.101 修复缓存问题。复盘还提到 Opus 4.7 在获得完整代码仓库上下文时，能够发现 4.6 版本遗漏的 bug。
English Summary: Anthropic's engineering team published a postmortem on recent Claude Code quality issues. A system prompt change adding length limits (≤25 words between tool calls, ≤100 words for final responses) to reduce verbosity, shipped with Opus 4.7 on April 16, unexpectedly degraded model intelligence. Additionally, a caching optimization incorrectly dropped prior reasoning from conversation history, affecting code review functionality. After user feedback, the team reverted Opus 4.7 default effort to xhigh on April 7 and fixed the caching bug in v2.1.101 on April 10. The postmortem noted Opus 4.7 could identify bugs missed by 4.6 when given complete repository context.
-
Scaling Managed Agents: Decoupling the brain from the hands（Anthropic Engineering）
中文摘要：Anthropic 工程团队发布《Scaling Managed Agents: Decoupling the brain from the hands》技术博客，介绍托管智能体系统的设计理念。该系统采用元架构（meta-harness）思路，通过通用接口将模型智能与具体执行工具解耦，使同一底层模型能够适配不同场景的智能体框架（如 Claude Code 或特定领域专用框架）。系统提供持久化会话存储、事件获取与转换、上下文管理等能力，支持长时运行任务。文章举例说明不同模型行为差异：Sonnet 4.5 曾出现"上下文焦虑"（接近上下文上限时过早结束任务），需通过架构层添加上下文重置解决，而 Opus 4.5 则无此问题。Managed Agents 作为托管服务，旨在通过稳定的接口抽象，适应未来不断演进的模型和框架实现。
English Summary: Anthropic's engineering team published a technical blog on "Scaling Managed Agents: Decoupling the brain from the hands," introducing the design philosophy of their managed agent system. The meta-harness approach decouples model intelligence from execution tools through general interfaces, allowing the same underlying model to adapt to different agent frameworks (like Claude Code or domain-specific harnesses). The system provides durable session storage, event fetching and transformation, and context management for long-horizon tasks. The article illustrates model behavioral differences: Sonnet 4.5 exhibited "context anxiety" (prematurely wrapping up tasks when approaching context limits) requiring harness-level context resets, while Opus 4.5 did not. Managed Agents aims to provide stable interface abstractions that outlast evolving model and framework implementations.
-
Microsoft says it has over 20M paid Copilot users, and they really are using it（TechCrunch AI）
中文摘要：微软宣布 Microsoft 365 Copilot 付费用户突破 2000 万，并强调用户活跃度真实增长。尽管外界长期质疑 Copilot 实际使用情况，微软表示该功能在 Word、Excel、Outlook 等应用中的使用量持续上升。Agent 模式成为增长驱动力，目前已作为 Copilot 及三大办公应用的默认体验。微软同时宣布支持多模型策略，用户可在聊天中默认访问多个模型，通过智能自动路由、批判与建议机制协同使用不同模型生成最优回复。微软强调 Copilot 不依赖单一模型（如 OpenAI），并已在平台中支持 Anthropic Claude 等其他模型。
English Summary: Microsoft announced that Microsoft 365 Copilot has surpassed 20 million paid users, emphasizing genuine engagement growth. Despite lingering skepticism about actual usage, Microsoft stated that usage within Word, Excel, and Outlook continues to rise. Agent mode is driving adoption and is now the default experience across Copilot and the three major Office apps. Microsoft also announced multi-model support, allowing users to access multiple models by default in chat, with intelligent auto-routing and critique-and-counsel mechanisms to combine models for optimal responses. The company emphasized that Copilot is not dependent on any single model like OpenAI, with Anthropic's Claude already supported on the platform.
-
Extracting contract insights with PwC’s AI-driven annotation on AWS（AWS ML Blog）
中文摘要：PwC与AWS联合发布AI驱动合同标注解决方案AIDA，利用Amazon Bedrock大语言模型和RAG技术实现合同智能分析。该系统支持模板化数据提取、单文档对话问答和跨文档全局搜索三大核心功能，可将合同审查时间缩短高达90%。架构上采用Amazon ECS、S3、RDS、OpenSearch Serverless等云原生服务，结合OCR、向量检索和LLM推理，为法律、合规和采购团队提供可溯源、可验证的合同洞察。某大型影视工作室应用后，版权研究时间减少90%，展示了该方案在媒体娱乐、房地产等行业的规模化应用潜力。
English Summary: PwC and AWS co-launched AIDA, an AI-driven contract annotation solution leveraging Amazon Bedrock LLMs and RAG to extract structured insights from legal documents. The system offers template-based extraction, document-level chat, and global cross-document search, reducing manual contract review time by up to 90%. Built on cloud-native AWS services including ECS, S3, RDS, and OpenSearch Serverless, AIDA combines OCR, vector retrieval, and LLM reasoning to provide traceable, verifiable contract intelligence for legal, compliance, and procurement teams. A major film studio achieved 90% reduction in rights research time, demonstrating scalability across media, entertainment, and real estate sectors.
-
Building the compute infrastructure for the Intelligence Age（OpenAI News）
中文摘要：OpenAI宣布其Stargate基础设施项目已提前完成原定2029年的10GW算力目标，过去90天新增超3GW容量。Stargate是OpenAI为构建通用人工智能所需算力基础而设立的长期项目，最新旗舰模型GPT-5.5即在得州Abilene数据中心训练完成。OpenAI强调算力是AI发展的核心输入，更多算力支持更强模型训练、更可靠的服务和更低的智能交付成本。公司采用合作伙伴模式推进，与Oracle、Vantage等合作建设数据中心，并承诺为当地社区创造就业、教育投资和负责任的水资源管理。
English Summary: OpenAI announced its Stargate infrastructure project has already surpassed its original 10GW compute target set for 2029, with over 3GW added in the past 90 days alone. Stargate is OpenAI's long-term initiative to build the compute foundation required for AGI, with its latest flagship model GPT-5.5 trained at the Abilene, Texas facility. The company emphasizes compute as the critical input enabling better model training, more reliable serving, and lower intelligence delivery costs over time. OpenAI pursues a partner-centric approach with Oracle, Vantage, and others for data center construction, while committing to local job creation, educational investments, and responsible water stewardship in host communities.
-
Presentation: Agents, Architecture, & Amnesia: Becoming AI-Native Without Losing Our Minds（InfoQ AI/ML）
中文摘要：InfoQ发布Tracy Bannon的演讲《Agents, Architecture & Amnesia》，以《魔法师的学徒》寓言警示无节制AI自主性的风险。演讲探讨从机器人到自主智能体的演进，指出盲目追求速度会导致"架构失忆症"——组织在快速采用AI的过程中丧失对系统设计和决策逻辑的追踪能力。Bannon强调在成为AI原生企业的过程中，必须保持架构治理和人工监督，避免过度自动化带来的不可控后果。该演讲为正在部署AI智能体的企业提供了关于自主性与可控性平衡的重要思考框架。
English Summary: InfoQ published Tracy Bannon's presentation "Agents, Architecture & Amnesia," using the Sorcerer's Apprentice fable to illustrate risks of unbridled AI autonomy. The talk explores the evolution from bots to autonomous agents, warning that reckless speed leads to "Architectural Amnesia"—where organizations lose track of system design and decision logic while rapidly adopting AI. Bannon emphasizes maintaining architectural governance and human oversight when becoming AI-native, avoiding uncontrollable consequences from excessive automation.
-
Cybersecurity in the Intelligence Age（OpenAI News）
中文摘要：OpenAI发布《智能时代的网络安全》行动计划，提出五大支柱应对AI驱动的网络威胁：普及AI网络防御工具、加强政府与行业协调、强化前沿网络安全能力管控、保持部署可见性与控制、赋能用户自我保护。OpenAI指出AI正在重塑网络安全格局，防御者和攻击者都在利用AI能力，因此需要与联邦和州政府及商业实体合作，通过民主制度和流程建立韧性，同时扩大可信主体获取防御技术的渠道。完整计划已以PDF形式公开发布。
English Summary: OpenAI published a "Cybersecurity in the Intelligence Age" action plan outlining five pillars to address AI-driven cyber threats: democratizing cyber defense, coordinating across government and industry, strengthening security around frontier cyber capabilities, preserving visibility and control in deployment, and enabling users to protect themselves. OpenAI notes AI is reshaping cybersecurity, with both defenders and attackers leveraging AI capabilities, necessitating collaboration with federal, state, and commercial entities. The plan emphasizes building resilience through democratic institutions while broadening access to defensive technologies for trusted actors. The complete plan is publicly available as a PDF.
-
[AINews] not much happened today（Latent Space）
中文摘要：Latent Space的AINews栏目承认当日AI新闻相对平淡，但汇总了值得关注的技术动态：vLLM 0.20发布带来TurboQuant 2-bit KV缓存、DeepSeek V4 MegaMoE支持等推理优化；Poolside开源33B MoE代码模型Laguna XS.2，可在单卡运行；NVIDIA发布30B多模态MoE模型Nemotron 3 Nano Omni，支持256K上下文和图文音视频理解，获主流平台同日上线；Mistral推出Workflows预览版，聚焦企业级智能体编排；本地离线智能体方案日趋成熟，Hugging Face、Gemma等推动端侧部署。此外，GPT-5.5 Pro在Epoch Capabilities Index达到159分，FrontierMath Tier 4解题率达40%。
English Summary: Latent Space's AINews acknowledged a quiet day in AI but highlighted notable developments: vLLM 0.20 released with TurboQuant 2-bit KV cache and DeepSeek V4 MegaMoE support for inference optimization; Poolside open-sourced Laguna XS.2, a 33B MoE coding model runnable on single GPU; NVIDIA launched Nemotron 3 Nano Omni, a 30B multimodal MoE with 256K context supporting text, image, video, and audio, with same-day availability across major platforms; Mistral introduced Workflows preview for enterprise agent orchestration; local offline agent solutions matured with Hugging Face and Gemma pushing on-device deployment. Additionally, GPT-5.5 Pro scored 159 on Epoch Capabilities Index with 40% on FrontierMath Tier 4.
-
[AINews] ImageGen is on the Path to AGI（Latent Space）
中文摘要：Latent Space 的 AINews 栏目探讨了 GPT-Image-2 在图像生成领域的持续爆发，认为高质量的图像生成能力是实现 AGI 的必要组成部分。文章指出，尽管各大实验室都在竞相模仿 Anthropic 专注于编程和企业 AI 的方向，但 GPT-Image-2 在创意应用、教育内容、流行文化和信息图表生成方面展现出独特价值。特别是当图像生成与 Codex 编码代理结合时，开发者可以在编码过程中实时生成所需素材，形成"闭环"工作流。文章还提到 Nano Banana、Grok Imagine 等模型的进展，强调多模态能力（语音和视觉生成）对于实现真正的通用人工智能至关重要。
English Summary: Latent Space's AINews discusses the continued explosion of GPT-Image-2 in image generation, arguing that high-quality image generation is a necessary component for achieving AGI. While labs race to emulate Anthropic's coding and enterprise AI focus, GPT-Image-2 demonstrates unique value in creative applications, educational content, pop culture, and infographic generation. When combined with Codex coding agents, developers can generate assets in real-time during coding, creating a "closed-loop" workflow. The article also covers progress on Nano Banana and Grok Imagine, emphasizing that multimodal capabilities (voice and visual generation) are essential for true artificial general intelligence.
-
Reading today's open-closed performance gap（Interconnects）
中文摘要：Nathan Lambert 在 Interconnects 博客中深入分析了开源与闭源模型之间的性能差距，指出单纯用一个数字来衡量这种差距会掩盖许多关键动态。文章讨论了影响评估结果的复杂因素，包括基准测试随时间的演变、模型实际性能与排名之间的关系，以及训练方法的变化。作者认为当前行业正处于以复杂编程和终端任务为重点的时代末期，前沿实验室正投入巨额资金掌握这些领域，同时开始向会计、法律、医疗等专业知识工作拓展。开源模型（尤其是中国实验室的模型）虽然在追赶，但在需要私有数据和复杂环境的领域可能难以跟上，因为前沿实验室通过购买新环境和数据集建立了类似芯片工厂的竞争优势。
English Summary: Nathan Lambert's Interconnects blog provides an in-depth analysis of the performance gap between open and closed models, arguing that reducing this gap to a single number obscures crucial dynamics. The article discusses complex factors affecting evaluation results, including benchmark evolution over time, the relationship between model performance and rankings, and changes in training methodologies. The author suggests the industry is at the end of an era focused on complex coding and terminal tasks, with frontier labs investing heavily while expanding into specialized domains like accounting, law, and healthcare. While open models (particularly from Chinese labs) are catching up, they may struggle in areas requiring private data and complex environments, as frontier labs build competitive advantages through acquiring new environments and datasets.
-
Building an emoji list generator with the GitHub Copilot CLI（GitHub AI/ML）
中文摘要：GitHub 博客分享了在 Rubber Duck Thursday 直播中使用 GitHub Copilot CLI 构建表情符号列表生成器的实践案例。该项目是一个终端应用，用户可以粘贴或输入项目列表，通过 AI 智能匹配相关表情符号，并将结果复制到剪贴板。开发过程采用了 Plan 模式进行需求规划和架构设计，然后使用 Claude Opus 4.7 实现代码。技术栈包括 OpenTUI 用于终端界面、GitHub Copilot SDK 提供 AI 能力、clipboardy 处理剪贴板功能。文章展示了 Copilot CLI 的多项特性，包括 Plan 模式、Autopilot 模式、多模型工作流、allow-all 工具标志以及 GitHub MCP 服务器的集成。
English Summary: GitHub Blog shares a practical case of building an emoji list generator using GitHub Copilot CLI during the Rubber Duck Thursday livestream. The project is a terminal application where users can paste or input a list of items, which AI intelligently matches with relevant emojis and copies the result to the clipboard. The development process used Plan mode for requirements planning and architecture design, then implemented code with Claude Opus 4.7. The tech stack includes OpenTUI for terminal UI, GitHub Copilot SDK for AI capabilities, and clipboardy for clipboard functionality. The article showcases several Copilot CLI features including Plan mode, Autopilot mode, multi-model workflows, the allow-all tools flag, and GitHub MCP server integration.
-
Build a personal organization command center with GitHub Copilot CLI（GitHub AI/ML）
中文摘要：GitHub 博客采访了工程师 Brittany Ellich，介绍她使用 GitHub Copilot CLI 构建的个人组织指挥中心项目。该项目旨在解决数字信息分散的问题，将分散在十几个不同应用中的内容统一到一个集中的空间中。Brittany 采用"先规划后实施"的工作流程，利用 AI 进行规划，使用 Copilot 进行实现，仅用一天时间就完成了第一个可用版本。她详细介绍了开发方法：Copilot 通过提问来明确需求，直到形成充分的实施计划，从而减少猜测并提高开发效率。她常用的工具栈包括 VS Code 的 Agent 模式进行同步开发，以及 Copilot Cloud Agent 进行异步开发。文章强调，从头开始构建解决方案从未如此简单，这是学习使用新 AI 工具的绝佳方式。
English Summary: GitHub Blog interviews engineer Brittany Ellich about her personal organization command center project built with GitHub Copilot CLI. The project aims to solve digital fragmentation by unifying content scattered across a dozen different apps into one centralized space. Brittany adopted a "plan-then-implement" workflow, using AI for planning and Copilot for implementation, completing the first working version in just one day. She details her development approach: Copilot interviews her with questions to clarify requirements until an adequate implementation plan is formed, reducing guesswork and improving efficiency. Her preferred tool stack includes VS Code's Agent mode for synchronous development and Copilot Cloud Agent for asynchronous tasks. The article emphasizes that building solutions from scratch has never been easier and is an excellent way to learn new AI tools.
-
Ollama is now powered by MLX on Apple Silicon in preview（Ollama Blog）
中文摘要：Ollama 官方博客宣布在 Apple Silicon 上推出基于 MLX 框架的预览版本，这是目前在苹果芯片上运行 Ollama 的最快方式。新版本利用苹果统一内存架构，在所有 Apple Silicon 设备上实现显著加速，在 M5、M5 Pro 和 M5 Max 芯片上更是利用新的 GPU 神经加速器来加速首 token 时间和生成速度。Ollama 0.19 还引入了对 NVIDIA NVFP4 格式的支持，在保持模型精度的同时减少内存带宽和存储需求，使本地运行结果与生产环境保持一致。此外，缓存系统得到升级，包括跨对话重用缓存、智能检查点存储和更智能的淘汰策略，使编码和代理任务更加高效。该版本特别针对 Qwen3.5-35B-A3B 模型进行了优化，适用于 Claude Code、OpenClaw 等编码代理场景。
English Summary: Ollama's official blog announces a preview version powered by Apple's MLX framework on Apple Silicon, representing the fastest way to run Ollama on Apple chips. The new version leverages Apple's unified memory architecture for significant acceleration across all Apple Silicon devices, with M5, M5 Pro, and M5 Max chips utilizing new GPU Neural Accelerators to speed up time-to-first-token and generation speed. Ollama 0.19 also introduces support for NVIDIA's NVFP4 format, maintaining model accuracy while reducing memory bandwidth and storage requirements, ensuring local results match production environments. Additionally, the caching system has been upgraded with cross-conversation cache reuse, intelligent checkpoint storage, and smarter eviction policies, making coding and agent tasks more efficient. This version is specifically optimized for the Qwen3.5-35B-A3B model, suitable for coding agents like Claude Code and OpenClaw.