九月二日

Claude says Fable 5.1 is now broadly available, while Mythos 5.1, a model for cyberdefenders and life scientists, is offered through trusted-access programs. Anthropic's Boris Cherny adds three concrete changes: cache reads fall from $1 to $0.25 per million tokens, making a typical Claude Code session up to 38% cheaper; biology safeguards intervene on benign requests 85% less often; and Claude Code users should encounter about 60% fewer cyber interventions per session. The release also targets a product-quality problem that benchmarks rarely capture: better writing and tone with less recognizable “Claude-speak.”

Claude 表示，Fable 5.1 现已全面开放；面向 cyberdefenders 与 life scientists 的 Mythos 5.1 则通过 trusted-access programs 提供。Anthropic 的 Boris Cherny 补充了三项具体变化：cache reads 从每百万 token 1 美元降至 0.25 美元，使典型 Claude Code session 的成本最多降低 38%；biology safeguards 对 benign requests 的干预减少 85%；Claude Code 用户每个 session 遇到的 cyber interventions 预计减少约 60%。这次发布还处理了一个 benchmark 很少衡量、却直接影响产品体验的问题：写作与语气更自然，并减少辨识度很高的 “Claude-speak”。

Box CEO Aaron Levie reports that Fable 5.1 improved seven percentage points over Fable 5 on Box's complex enterprise-work evaluation for unstructured-data tasks. The gains were uneven but revealing: 17% in a financial-services profit projection that required applying capital allowances before tax, 37% in a technology cost analysis where the model had to recognize an ambiguous normalization, and 16% in a public-sector weighted-mean task where one classification error could corrupt the full ranking. Levie says the model will soon enter Box AI Studio, turning the benchmark improvement into an option for custom agents working with enterprise content.

Box CEO Aaron Levie 表示，在 Box 针对复杂 enterprise work 的 unstructured-data evaluation 中，Fable 5.1 比 Fable 5 提升了 7 个百分点。不同任务的提升幅度并不一致，却很能说明问题：在必须先应用 capital allowances、再计算 tax 的金融利润预测中提升 17%；在 normalization 定义含糊、需要模型主动识别歧义的技术成本分析中提升 37%；在一个分类错误就可能破坏完整排名的 public-sector weighted-mean 任务中提升 16%。Levie 表示，该模型将很快进入 Box AI Studio，把 benchmark 的提升转化为企业基于自身内容构建 custom agents 的实际选项。

Anthropic introduced Enterprise Frontier Safeguards, or EFS, promising enterprise privacy equivalent to zero data retention while adding protection against adversarial use; rollout begins in phases this fall. Anthropic researcher Alex Albert describes the design as “ZDR++”: company data remains in the company's own cloud, while an automated monitoring layer detects risky patterns across agent activity and flags them to the enterprise team. As agents gain access to more internal systems, he expects this combination of privacy, observability, and risk mitigation to become a standard enterprise requirement.

Anthropic 推出了 Enterprise Frontier Safeguards，也就是 EFS，在提供等同于 zero data retention 的企业隐私保障同时，增加对 adversarial use 的防护；该能力将从今年秋季开始分阶段上线。Anthropic researcher Alex Albert 把它称为 “ZDR++”：企业数据留在自己的 cloud 中，同时由 automated monitoring layer 跨 agent activity 识别风险模式并提醒企业团队。随着 agents 获得越来越多 internal systems 的访问权限，他预计这种把 privacy、observability 与 risk mitigation 结合起来的方案，会成为企业部署 AI 的标准要求。

OpenAI CEO Sam Altman says the company will launch its next model soon and that Astra finished training some time ago, representing a significant step in both capability and alignment. He also says OpenAI has slowed work on models after Astra when necessary to allow more safety and alignment work. His central claim is that no one fully understands the consequences of increasingly capable AI, so models should reach the real world through an iterative loop in which society and the technology adapt together, even while the company lives with the tension between excitement and caution.

OpenAI CEO Sam Altman 表示，公司将很快发布下一款 model；Astra 已在一段时间前完成训练，并在 capability 与 alignment 两方面实现显著进步。他同时指出，对于 Astra 之后的 models，OpenAI 会在必要时放慢进度，以便投入足够的 safety 与 alignment 工作。他的核心判断是，没人能够完全理解能力快速增强的 AI 会带来什么后果，因此 models 应通过一种 iterative loop 进入现实世界，让社会与技术共同适应，即使公司自身也持续处于兴奋与谨慎并存的张力之中。

Meta Senior Director of AI Madhu Guru argues that a self-improving product requires five explicit foundations: primary, secondary, and guardrail metrics; a strategy and roadmap tied to the metrics that matter most; a knowledge base of prior product decisions and principles; connections to dashboards, APIs, MCPs, and other internal tools; and a harness that understands the full product-development flow. He separately points to enterprise-owned post-training systems, strong evals, and a data flywheel as a major opportunity, using Shopify's machine-learning team as the example. The implication is that improvement becomes repeatable only when an agent can observe, judge, and act inside a governed product system.

Meta Senior Director of AI Madhu Guru 认为，self-improving product 需要五项明确基础：primary、secondary 与 guardrail metrics；与最重要 metrics 对齐的 strategy 和 roadmap；记录过往产品决策与原则的 knowledge base；连接 dashboards、APIs、MCPs 及其他 internal tools；以及理解完整 product-development flow 的 harness。他还指出，企业自建 post-training systems、强 evals 与 data flywheel 是一个巨大机会，并以 Shopify 的 machine-learning team 为例。这里的含义是，只有当 agent 能在一个受治理的产品系统内进行观察、判断与行动时，持续改进才可能成为可重复的能力。

Incoming OpenAI product staff member Nan Yu argues that making agents “less annoying” has substantial product value because users who rage-quit never reach the underlying capability. He sees an untapped role for UX designers as conversation and rhetoric designers, and points to Fable 5.1's less formulaic writing as an example of a seemingly small change that materially alters usage. This shifts evaluation beyond whether an agent completes a task toward whether its language keeps the user engaged long enough to receive the result.

即将加入 OpenAI product staff 的 Nan Yu 认为，让 agents “less annoying” 本身就有很大的产品价值，因为一旦用户 rage-quit，再强的底层能力也无法兑现。他看到一个尚未被充分开发的 UX 方向：让设计师成为 conversation 与 rhetoric designers；他把 Fable 5.1 更少套路化的写作视为例子，说明看似细小的变化也会显著改变实际使用体验。这让 evaluation 不再只看 agent 能否完成任务，也要看它的语言是否能让用户持续参与，直到真正获得结果。

AI educator Peter Yang recommends keeping only a small set of actively used skills, shortening their instructions, and running Fable 5.1's `/claude-api prompt-audit` to find redundant rules. He also identifies a harder maintenance failure: after a near miss, asking AI to rewrite a skill for a one-shot success can overfit it to one conversation and cause gradual drift. Anthropic's Thariq offers a related runtime optimization for Fable 5.1: use lower effort on tasks with fewer edge cases or less need for verification, now that switching effort no longer breaks the prompt cache.

AI educator Peter Yang 建议只保留少量真正会使用的 skills，尽量缩短 instructions，并运行 Fable 5.1 的 `/claude-api prompt-audit` 查找冗余规则。他还指出一个更棘手的维护问题：当一次执行没有完全成功后，让 AI 为了下次 one-shot 成功而重写 skill，很容易对单次 conversation 产生 overfitting，并让 skill 逐渐 drift。Anthropic 的 Thariq 则给出一个相关的 Fable 5.1 runtime 优化：对于 edge cases 较少、verification 要求较低的任务，可以使用 lower effort，因为现在切换 effort 已不会破坏 prompt cache。

Vercel CEO Guillermo Rauch says Fluid underpins the company's build performance, sandbox reliability and concurrency, and 30-minute Function durations. His goal is one compute platform with shared Dockerfiles, security perimeters, networking, and file systems across products, including the prospect of applying Sandbox egress-firewall protections against agent escapes to Builds, Functions, and Servers. At the framework layer, Vercel is also partnering with Tanner Linsley and committing service and support to customers whether they choose Next.js or TanStack, broadening its platform posture beyond a single React framework.

Vercel CEO Guillermo Rauch 表示，Fluid 支撑了公司的 build performance、sandbox reliability 与 concurrency，以及 30 分钟的 Function durations。他的目标是打造一个统一 compute platform，让不同产品共享 Dockerfiles、security perimeters、networking 与 file systems，并有望把 Sandbox 中防止 agent escapes 的 egress-firewall protections 应用到 Builds、Functions 和 Servers。在 framework 层，Vercel 也正与 Tanner Linsley 合作，并承诺无论客户选择 Next.js 还是 TanStack，都提供相应的服务与支持，使平台定位不再局限于单一 React framework。

FPV Ventures partner Nikunj Kothari demonstrates WebMCP as a way for websites to expose native tool calls, including UI controls and interactive elements that agents can use to build their own views. In his El Niño tracker demo, an agent creates a view, preserves human edits, and produces a shareable link for other agents or people. The practical shift is from agents visually approximating a site to sites deliberately publishing an interaction surface for them.

FPV Ventures partner Nikunj Kothari 演示了 WebMCP 如何让网站暴露 native tool calls，其中包括 agents 可用来构建自定义视图的 UI controls 与 interactive elements。在他的 El Niño tracker demo 中，agent 能创建视图、保留人工编辑，并生成可分享给其他 agents 或用户的链接。这里的实际转变，是从 agent 通过视觉方式近似操作网站，走向网站主动为 agent 发布明确的 interaction surface。

Box CEO Aaron Levie expects AI for cybersecurity to accelerate sharply as models become much better at finding and exploiting vulnerabilities, with frontier models leading but open weights close behind. Because enterprises already face more findings than teams can comfortably process, he argues that AI-assisted triage and automated fixes, kept under human oversight, are becoming the only workable response. The likely labor effect is not the disappearance of security work but a larger need for people who can supervise and harden these systems.

Box CEO Aaron Levie 预计，随着 models 在发现和利用 vulnerabilities 方面快速增强，AI for cybersecurity 将显著加速；frontier models 仍然领先，但 open weights 已紧随其后。企业本就面对超出团队处理能力的大量 findings，因此他认为，在 human oversight 下使用 AI 辅助 triage 并自动修复，正成为唯一可行的应对方式。对劳动力的影响更可能不是 security work 消失，而是市场更需要能够监督并加固这些系统的人。

The Takeaway: Peregrine's Nick Noone and Ben Rudolph argue that useful AI for cities begins with governed access to data that public institutions already own, then combines that foundation with forward-deployed teams and long-running agents that can reach verifiable operational outcomes.

核心结论：Peregrine 的 Nick Noone 与 Ben Rudolph 认为，真正对城市有用的 AI，起点是以受治理的方式访问公共机构已经拥有的数据，再把这一基础与 forward-deployed teams、能够长期运行的 agents 结合，最终交付可验证的 operational outcomes。

Peregrine inverts the collection-centered model used by many public-safety vendors. Instead of installing more systems to gather more information, it joins disparate data already owned by each institution, while treating permissions, governance, sovereignty, and selective sharing as core product infrastructure. Noone's forward-deployed philosophy is equally specific: engineers should temporarily co-own the customer's problem, suspend their ego about decades of institutional context, and get to the customer's outcome three to five times faster without claiming the customer's win as their own.

Peregrine 反转了许多 public-safety vendors 以数据采集为中心的模式。它不是安装更多系统来收集更多信息，而是连接各机构本就拥有的 disparate data，并把 permissions、governance、sovereignty 与 selective sharing 当作核心产品基础设施。Noone 的 forward-deployed 哲学同样具体：工程师应在一段时间内与客户共同承担问题，面对几十年的 institutional context 时放下自身的 ego，并以三到五倍的速度帮助客户实现 outcome，同时不把客户的成果据为己有。

Rudolph says “95% of the work is what happens before the user types in the question.” Peregrine puts roughly half its engineers on the data platform and has integrated tens of thousands of customer datasets. Agents now write about 90% of the Python-notebook work used for integrations, under deployment-team oversight; they can run for hours, inspect databases and ontologies, split work among subagents, and return to an orchestrator. This preparation supports higher-level use cases, including connecting weather, terrain, incident, and 911 data to explain a Florida county's unusual surge of more than 100 water rescues.

Rudolph 表示：“95% of the work is what happens before the user types in the question.” Peregrine 约有一半工程师投入 data platform，并已整合数以万计的客户 datasets。如今，agents 会在 deployment team 的监督下完成约 90% 基于 Python notebook 的 integration 工作；它们可以连续运行数小时，检查 databases 与 ontologies，把任务拆给 subagents，再汇总给 orchestrator。这些前置准备支撑了更高层的 use cases，例如连接 weather、terrain、incident 与 911 data，解释 Florida 某 county 为何异常出现超过 100 次 water rescues。

The same architecture powers a cold-case agent built from 200 to 300 gigabytes of video, audio, images, PDFs, and records. Peregrine first tested whether the agent could reproduce detectives' reasoning in a case that ended in an exoneration; in a later Wisconsin case, it found a small number of phone-location records inside roughly 300 gigabytes of evidence that placed a suspect both at the crime scene and where the body was found. The founders frame these systems as institutional infrastructure, not centralized intelligence: each jurisdiction keeps ownership and control, while forward-deployed work supplies both customer outcomes and the highest-quality signal for the next reusable product primitive.

同一套架构也支撑了 cold-case agent，它要处理 200 到 300GB 的视频、音频、图片、PDF 与记录。Peregrine 最初测试 agent 能否重现 detectives 在一起最终洗清错误定罪者罪名的案件中的推理；之后在 Wisconsin 的另一起案件里，它从约 300GB 证据中找到了少量 phone-location records，把嫌疑人同时定位在 crime scene 与发现遗体的地点。两位创始人把这类系统定义为 institutional infrastructure，而不是 centralized intelligence：每个 jurisdiction 保留 ownership 与 control，forward-deployed work 则同时交付客户结果，并为下一项可复用 product primitive 提供最高质量的信号。