让变化经过筛选，
再抵达你。

AWS 宣布 Claude Fable 5.1 可通过 Amazon Bedrock 与 Claude Platform on AWS 使用，面向编码、科学研究和长时企业任务。Anthropic 将它列为 Covered Model；Bedrock 默认要求 aws_review 模式，提示和输出最多保留 30 天供 AWS 安全复核。符合 Enterprise Frontier Safeguards 条件的内部用途可选择零数据保留。能力改进来自 Anthropic 与 AWS 的报告。

AWS has made Claude Fable 5.1 available through Amazon Bedrock and Claude Platform on AWS for coding, scientific research, and long-running enterprise work. Anthropic classifies it as a Covered Model: Bedrock normally requires aws_review mode, retaining prompts and outputs for up to 30 days for AWS safety review. Eligible internal use under Enterprise Frontier Safeguards can use zero data retention. Capability improvements are reported by Anthropic and AWS.

OpenAI 发布 Epic 电子健康记录集成与 Healthcare Public Data 插件：授权机构可在 ChatGPT 中审阅患者上下文，并查询 PubMed、DailyMed、ClinicalTrials.gov 和 CMS Coverage 等九类官方来源。公司称医生在 27 个临床用例的 4,363 次评分中，将 99.1% 的响应评为安全；这是 OpenAI 的产品评估，不代表独立临床结局研究，输出仍需专业人员复核。

OpenAI introduced an Epic electronic-health-record integration and a Healthcare Public Data plugin, letting authorized organizations review patient context and query nine official sources including PubMed, DailyMed, ClinicalTrials.gov, and CMS Coverage. The company reports that physicians rated 99.1% of responses safe across 4,363 ratings and 27 clinical use cases. This is an OpenAI product evaluation, not an independent patient-outcomes study, and outputs still require professional review.

OpenAI 汇总 Basis、Clay 与 Exa Labs 的案例，分别把智能体用于员工入职、账户管理和开发者集成。文章同时引用 Enterprise Signals：AI 使用量前 10% 的企业，每位活跃用户输出 token 数是典型企业的 8.3 倍，1 月时为 2.6 倍。该指标和案例来自 OpenAI 及客户观察，说明采用深度，不直接证明生产率或投资回报。

OpenAI profiles Basis, Clay, and Exa Labs using agents for employee onboarding, account management, and developer integrations. It also cites Enterprise Signals: firms in the top 10% of AI usage now generate 8.3 times as many output tokens per active user as typical firms, up from 2.6 times in January. These OpenAI and customer observations measure usage depth, not productivity or return on investment by themselves.

Jamf 与 AWS 公开一套生产架构：按工程师统计每日 Bedrock 成本，在预算达到 80% 时限制高成本 Opus、达到 100% 时进一步限制 Sonnet，同时保留低成本 Haiku。Athena 计算花费，Lambda 每 15 分钟检查，IAM 策略在无需重新登录的情况下生效，并提供限时例外。这是单一客户方案，阈值与收益需按组织重新设计。

Jamf and AWS describe a production architecture that tracks daily Bedrock cost per engineer, restricts the higher-cost Opus model at 80% of budget, further restricts Sonnet at 100%, and keeps the lower-cost Haiku option available. Athena calculates spend, Lambda checks every 15 minutes, and IAM policies apply without reauthentication, with time-boxed exceptions. This is one customer pattern; thresholds and benefits need redesign for each organization.

AWS 发布从概念验证走向生产的 Amazon Quick 安全实践，以 5,000 名员工、五个部门和三类权限受众的示例说明四种模式：按授权塑形数据集、隔离智能体、分类文档、为外部动作设置人工审批。文章是 AnyCompany 场景的技术指南，不是一次现实部署成效评估，但清楚展示了仅依赖单层权限为何不足。

AWS published production-security guidance for Amazon Quick using a sample organization with 5,000 employees, five departments, and three audience levels. It demonstrates four patterns: shaping datasets around authorization, isolating agents, classifying documents, and placing human approval gates on outbound actions. This is an AnyCompany technical walkthrough rather than a deployment-outcomes study, but it shows why a single permission layer is insufficient.

同行评议研究发布开源 AdaptiveFlow，为 690 亿个可对接分子提供筛选就绪数据，并整合超过 1,500 种对接协议。作者报告，自适应预筛在部分设置中可将计算成本降低至多 1,000 倍，并通过生化实验与晶体结构验证了两个靶点的纳摩尔级抑制剂。结果支持早期先导发现，不代表候选物已经证明安全、有效或可成为药物。

A peer-reviewed study introduces the open-source AdaptiveFlow platform with a screening-ready library of 69 billion dockable molecules and more than 1,500 docking protocols. The authors report that adaptive prescreening can reduce computational cost by up to 1,000-fold in some settings, and they validated nanomolar inhibitors for two targets through biochemical assays and crystal structures. This supports early hit discovery; it does not show that candidates are safe, effective, or ready to become medicines.