九月二十三日

Anthropic’s Cat Wu says Claude Opus 5.5 is now the default in Claude Code and the Claude app, including Cowork, for Pro, Max, and Team plans. Products default to medium effort, which she describes as comparable to Fable 5.1 in intelligence but faster; she also says rate limits go 25% further than with Opus 5. Her own experience emphasizes clearer communication and better adaptation to her writing style, while the Claude account separately confirms the model’s availability. The concrete product change combines a new default model, a default effort setting, and more work within existing usage limits.

Anthropic 的 Cat Wu 表示，Claude Opus 5.5 已成为 Pro、Max 和 Team 套餐在 Claude Code 及 Claude 应用中的默认模型，覆盖 Cowork。各产品默认使用 medium 推理强度；她称，这一设置的智能水平与 Fable 5.1 相当，但速度更快，同样的使用限额也能比 Opus 5 多支撑 25% 的用量。她的个人体验主要体现在沟通更清楚、更加贴合自己的写作风格，Claude 官方账号也另行确认了模型已可使用。这次产品调整同时涉及默认模型、默认推理强度，以及现有额度能够完成的工作量。 [Cat Wu’s announcement](https://x.com/_catwu/status/2102437713781944397) · [Claude availability](https://x.com/claudeai/status/2102471892099866883)

OpenAI’s Thibault Sottiaux announces GPT-6 Sol and Luna, describing broad capability improvements that also extend to writing and overall response quality. He says API pricing is being permanently reduced by 50%, with usage also going further on subscriptions, and that all Plus, Pro, and Business accounts are receiving a banked reset. His accompanying explanation ties the releases to a focus on efficiency and making intelligence more widely available, using stronger top-end models to improve the rest of the lineup. The announcement supplies a relative price reduction, but no absolute token prices or detailed benchmark results.

OpenAI 的 Thibault Sottiaux 宣布推出 GPT-6 Sol 和 Luna，称模型能力得到全面提升，写作和整体回答质量也有所改善。他表示，API 价格将永久下调 50%，订阅额度也能支持更多使用，所有 Plus、Pro 和 Business 账户还将获得一次可留待使用的额度重置。他在相关帖子中把这些发布与效率提升、扩大智能能力的可及性联系起来，强调借助最强模型来带动其他模型进步。公告给出了相对降价幅度，但没有提供具体的 token 单价或详细基准测试结果。 [Launch and pricing](https://x.com/thsottiaux/status/2102463847714247142) · [Efficiency strategy](https://x.com/thsottiaux/status/2102440619616682120) · [Link shared with the launch](https://t.co/00DRh1sRrO)

Box CEO Aaron Levie reports that internal Box Agent tests on complex enterprise work with unstructured data used 63% fewer tokens, produced 42% less verbose responses, and ran 30% faster with Opus 5.5 than with Opus 5. Reported task-accuracy improvements included 39% for financial due diligence and 65% for cloud-cost analysis: the former used 82% fewer tokens while checking a year of transactions for pricing-tool errors, and the latter used 70% fewer tokens while maintaining the right calculation basis and source-data units. These are Box’s reported results on its own tasks, and Levie says access through Box AI Studio will follow shortly. In a separate post, he argues that falling task costs expand the feasible scope of agents, from security reviews and log analysis to multi-agent workflows, potentially increasing overall AI use.

Box CEO Aaron Levie 表示，在 Box Agent 处理非结构化数据、执行复杂企业任务的内部测试中，Opus 5.5 相比 Opus 5 少用了 63% 的 token，回答冗长程度降低 42%，速度提升 30%。他报告的任务准确率提升包括财务尽调的 39% 和云成本分析的 65%：前者检查一年的交易记录以找出定价工具中的计算错误，token 用量减少 82%；后者正确处理了计算口径和源数据的单位，token 用量减少 70%。这些数字来自 Box 对自身任务的测试，Levie 表示 Box AI Studio 很快也将支持该模型。他在另一条帖子中提出，单项任务成本下降，会让安全审查、日志分析和多 agent 工作流等更多场景具备部署条件，从而可能扩大 AI 的总使用量。 [Box task results](https://x.com/levie/status/2102448415775051790) · [Levie’s cost-and-demand argument](https://x.com/levie/status/2102477253070430322)

Anthropic’s Boris Cherny reports using Opus 5.5 and Lean to formally verify the Claude Agent SDK, with a couple of short prompts leading to 16 pull requests fixing bugs and race conditions. He also uses TLA+, sometimes combining the two to investigate data flow, concurrency, and state management, despite not knowing either language well himself. In a separate migration experiment, he says Opus 5.5 and Fable 5.1 each ported HAProxy from C to Rust and passed nearly all of HAProxy’s tests, with Opus finishing in 9.5 hours versus 12 hours at 51% lower cost. The examples cover both bug discovery and a substantial code migration, but “nearly all” passing tests does not establish complete equivalence of the ports.

Anthropic 的 Boris Cherny 表示，他让 Opus 5.5 使用 Lean 对 Claude Agent SDK 进行形式化验证，只用了几个简短的 prompt，就产生了 16 个修复 bug 和竞态条件的 PR。他还会使用 TLA+，有时将两者结合起来检查数据流、并发和状态管理问题，尽管自己并不熟悉这两种语言。在另一项代码迁移实验中，他称 Opus 5.5 和 Fable 5.1 都将 HAProxy 从 C 移植到了 Rust，并通过了几乎全部 HAProxy 测试；Opus 用时 9.5 小时，Fable 用时 12 小时，前者成本低了 51%。这些案例同时涉及缺陷发现和较大规模的代码迁移，但“几乎全部”测试通过，并不能证明移植结果与原实现完全等价。 [Formal verification report](https://x.com/bcherny/status/2102543349102338309) · [HAProxy migration comparison](https://x.com/bcherny/status/2102439069053747549) · [Link shared with the comparison](https://t.co/GJ9FaS3Erx)

Vercel CEO Guillermo Rauch reports fresh Next.js eval results of 97% for Opus 5.5, GPT 6 Sol, and Fable 5.1, versus 94% for Grok 4.7. He says Grok is 2–7 times cheaper, pairing a small score gap in this evaluation with a substantial reported cost difference. The post provides neither the task breakdown nor the pricing basis, so the numbers support a comparison within the reported Next.js evaluation rather than a general ranking of coding ability.

Vercel CEO Guillermo Rauch 公布了新一轮 Next.js 评测结果：Opus 5.5、GPT 6 Sol 和 Fable 5.1 都是 97%，Grok 4.7 为 94%。他同时称 Grok 便宜 2–7 倍，在这项评测中，较小的分数差距伴随着明显的成本差距。帖子没有列出具体任务构成或价格比较口径，因此这些数字适合用于理解这次 Next.js 评测中的表现，不能直接当作通用编程能力排名。 [Next.js eval results](https://x.com/rauchg/status/2102519097770885231) · [Linked evaluation context](https://t.co/BAUg14981G)

OpenClaw and OpenAI’s Peter Steinberger says intermittent ChatGPT crashes after updating to macOS 27 led to Astra finding an approximately 14-year-old bug in libuv. This is a concrete debugging report reaching into a dependency, although the post does not describe the root cause, provide a patch, or establish whether a fix has shipped.

OpenClaw 和 OpenAI 的 Peter Steinberger 表示，升级 macOS 27 后，ChatGPT 偶尔会崩溃，Astra 在排查过程中找到了 libuv 中一个存在约 14 年的 bug。这是一个追查到依赖库层面的具体调试案例，不过帖子没有说明根因、提供补丁，也没有确认修复是否已经发布。 [Debugging report](https://x.com/steipete/status/2102501642176528743) · [Attached reference](https://t.co/JE8GssiRXJ)

Anthropic researcher Alex Albert shares an Opus 5.5 demonstration of San Francisco’s Market Street before the 1906 earthquake, along with the prompt used in Claude Tag. The prompt asks the model to first assemble building footprints, heights, materials, occupants, sources, and confidence levels from historical maps, film, photographs, and topography, then generate the scene in Blender Python without downloaded meshes, textures, or HDRIs. It specifies reusable generators for buildings and street objects and a ten-second video, with each building traceable to the source file. The useful engineering detail is the explicit separation of evidence gathering from procedural construction; historical accuracy is Albert’s description of the demo, not something established by the prompt alone.

Anthropic 研究员 Alex Albert 展示了 Opus 5.5 重建 1906 年地震前 San Francisco 的 Market Street 的案例，并分享了在 Claude Tag 中使用的 prompt。这个 prompt 要求模型先从历史地图、影像、照片和地形资料中整理建筑轮廓、高度、材质、使用者、来源和置信度，再用 Blender Python 生成场景，不使用下载的网格、纹理或 HDRI。它还要求为建筑和街道物件编写可复用的生成器，制作一段十秒视频，并让每栋建筑都能追溯到源数据文件。值得借鉴的工程细节，是将证据收集与程序化建模明确分开；“符合历史”是 Albert 对演示的描述，仅凭 prompt 本身还不能证实。 [Demo](https://x.com/alexalbert__/status/2102466523164274839) · [Full prompt](https://x.com/alexalbert__/status/2102466524934271381) · [Attached visual](https://t.co/aisvFLFGy3)

Anthropic’s Thariq argues that better model capabilities should buy more time for understanding users, running experiments, building prototypes, and learning unfamiliar subjects, rather than simply multiplying the number of features shipped. His game-development example makes the distinction concrete: 3D generation helps bring an imagined game to life, but a satisfying core game loop still comes first. He also says workflows are a large part of his Claude usage and welcomes intelligence comparable to Fable at a cost that makes those workflows practical.

Anthropic 的 Thariq 认为，模型能力提升后，应把省下来的时间用于理解用户、开展实验、制作原型和学习不熟悉的领域，而不是简单地把上线功能的数量翻倍。他用游戏开发说明了这一点：3D 生成可以帮助想象中的游戏变成可见的作品，但首先仍要做出让人愿意反复玩的核心循环。他还表示，工作流已占据自己使用 Claude 的很大一部分，因此很看重以可承受的成本获得接近 Fable 的智能水平，让这些工作流真正用得起来。 [Product-development argument](https://x.com/trq212/status/2102548686303854790) · [Game-loop example](https://x.com/trq212/status/2102549030303867257) · [Workflow economics](https://x.com/trq212/status/2102477527688388752)

SPC General Partner Aditya Agarwal points to autonomous vehicles as an earlier arena for AI safety debates after hosting a conversation with Waymo’s Dmitri Dolgov. What stood out to him was Waymo’s extensive evaluation and testing infrastructure, which supports confidence in deploying two-ton vehicles at 30 mph in urban environments. His post emphasizes the testing behind real-world deployment, without supplying implementation details about that infrastructure.

SPC General Partner Aditya Agarwal 在主持与 Waymo 的 Dmitri Dolgov 的交流后，指出自动驾驶早已是 AI 安全讨论的重要场景。他最关注的是 Waymo 大量投入建设的评估与测试基础设施：正是这些工作，让团队有信心将重达两吨、以每小时 30 英里行驶的车辆部署到城市环境中。他的帖子强调了真实世界部署背后的测试体系，但没有披露这套基础设施的具体实现。 [Waymo evaluation perspective](https://x.com/adityaag/status/2102457464432284019)