🤖 核心模型矩阵🤖 Model Matrix
Claude 5 系列 (Anthropic) vs Gemini (Google) — 2026 年两大旗舰 AI 平台全面对比。本页最后核实:2026 年 10 月 7 日。三件事记住就行:日常用 Sonnet 5.5(9/28 发布,取代 Sonnet 5)、难任务用 Opus 5.5(9/22 发布,取代 Opus 5)、Fable 5.1 基本用不上(Opus 5.5 在官方公布的基准上已超过它,价格只要四成)。
Claude 5 family (Anthropic) vs Gemini (Google) — a full comparison of 2026's two flagship AI platforms. Last fact-checked: October 7, 2026. Three things to remember: Sonnet 5.5 for daily work (released 9/28, replaces Sonnet 5), Opus 5.5 for hard work (released 9/22, replaces Opus 5), and you'll rarely need Fable 5.1 (Opus 5.5 beats it on Anthropic's published benchmarks at 40% of the price).
Opus 系列的最新一代,比 Opus 5 更强,反而更便宜。在 Anthropic 公布的编码、agent、知识工作、电脑操作几项基准上,它都超过了原旗舰 Fable 5.1,价格却只有 Fable 5.1 的四成——这是本页选型口径改变的主要原因。所有付费套餐(Pro / Max / Team / 企业版)都能用。站内以前写 "Opus 5" 的地方,现在都指 Opus 5.5。
The newest Opus is stronger than Opus 5, and cheaper. On the coding, agent, knowledge-work, and computer-use benchmarks Anthropic published, it beats the former flagship Fable 5.1 on every one, at 40% of Fable 5.1's price — the main reason this page's model advice has changed. Available on every paid plan (Pro / Max / Team / Enterprise). Wherever this site used to say "Opus 5," read Opus 5.5.
| 维度Dimension | 详情Detail |
|---|---|
| 定位Positioning | 长时间运行的 agent 编码与知识工作,需要判断稳、想得细的复杂任务。比 Opus 5 快约 30%;做 agent 编码时调用次数少约 40%、token 用量减半。Long-running agentic coding and knowledge work — complex jobs that need careful judgment. About 30% faster than Opus 5, and on agentic coding it makes ~40% fewer calls using half the tokens. |
| 价格Pricing | 输入 $4 / 百万 token,输出 $20 / 百万 token(Opus 5 为 $5/$25);缓存读取 $0.20(Opus 5 为 $0.50)。官方口径:常规任务比 Opus 5 省约 40%。fast mode:$8/$40。$4 / M input, $20 / M output (Opus 5 was $5/$25); cache reads $0.20 (Opus 5: $0.50). Anthropic puts it at ~40% cheaper than Opus 5 on typical workloads. Fast mode: $8/$40. |
| 上下文 & 思考Context & thinking | 100 万 token 上下文,单次最多 12.8 万 token 输出。自适应思考常开、不能关,用"努力档"控制想多深(API 默认 medium)。1M-token context, up to 128K tokens of output per request. Adaptive thinking is always on and can't be disabled; the effort setting controls how deep it goes (API default: medium). |
| 实测表现Benchmarks | 括号内依次为 Fable 5.1 / Opus 5:Terminal-Bench 4.0 66.4%(55.8% / 52.3%)、CursorBench 4.0 57.8%(51.8% / 46.6%)、知识工作 GDPval-AA 1846(1735 / 1708)、电脑操作 OSWorld 2.1 81.8%(80.7% / 74.0%)。Figures in brackets are Fable 5.1 / Opus 5: Terminal-Bench 4.0 66.4% (55.8% / 52.3%), CursorBench 4.0 57.8% (51.8% / 46.6%), knowledge-work GDPval-AA 1846 (1735 / 1708), computer-use OSWorld 2.1 81.8% (80.7% / 74.0%). |
| 安全Safety | 更能抵抗提示词注入(网页或文件里藏指令骗 AI);试图绕过限制的情况比 Opus 5 少 85%;生物与网络安全防护与 Fable 5.1 同级。Stronger resistance to prompt injection (instructions hidden in a web page or file to trick the AI); tries to work around its boundaries 85% less often than Opus 5; biology and cyber safeguards match Fable 5.1. |
| 获取方式Availability | claude.ai(所有付费套餐,官方同时上调了 Pro / Max / Team / 企业版的 5 小时用量上限)、Claude Code(已是默认的 Opus)、Claude Cowork、API,以及 Amazon Bedrock / Google Cloud / Microsoft Foundry。Claude Code 里用 /model claude-opus-5-5 切换。claude.ai (every paid plan — and Anthropic raised the five-hour usage limits on Pro, Max, Team, and Enterprise at the same time), Claude Code (now the default Opus), Claude Cowork, the API, plus Amazon Bedrock / Google Cloud / Microsoft Foundry. Switch in Claude Code with /model claude-opus-5-5. |
| VWD 怎么用VWD takeaway | 难任务的首选:复杂型材受力计算、合同与规范合规审查、跨部门流程梳理、完整代码编写与审查。以前"值得开 Fable"的大活(跑几个小时的 agent、大型迁移),现在也先用 Opus 5.5。日常任务仍用 Sonnet 5.5(更快、更便宜、够用)。The first pick for hard work: complex structural calcs, contract and code-compliance review, untangling cross-department processes, full code generation and review. The big jobs that used to justify Fable (multi-hour agents, large migrations) start on Opus 5.5 now too. Stay on Sonnet 5.5 for everyday work — faster, cheaper, and good enough. |
官方公告:anthropic.com/claude-opus-5-5 · 📘 官方提示词指南:Prompting Claude Opus 5.5(英文,偏 API / 开发)
Official announcement: anthropic.com/claude-opus-5-5 · 📘 Official prompting guide: Prompting Claude Opus 5.5
Opus 5.5 发布不到一周后推出的同代 Sonnet,定位是又快又省的日常搭档。官方说它最擅长边界清楚的日常任务、修 bug,以及做出像样的文档、幻灯片和表格——正好是 VWD 大多数人每天在做的事。价格与 Sonnet 5 相同,但快 30% 以上,每个任务最多便宜 30%。
The same-generation Sonnet, released less than a week after Opus 5.5 and pitched as a faster, cheaper everyday partner. Anthropic says it is strongest at well-scoped everyday tasks, fixing bugs, and producing polished documents, slides, and spreadsheets — which is most of what people at VWD do all day. Same price as Sonnet 5, but 30%+ faster and up to 30% cheaper per task.
| 维度Dimension | 详情Detail |
|---|---|
| 定位Positioning | 速度与智能的最佳组合。知识工作基准 GDPval-AA 1844 分,与 Opus 5.5(1846)几乎打平;Terminal-Bench 4.0 70.6%,甚至略高于 Opus 5.5(66.4%)。The best balance of speed and intelligence. It scores 1844 on the GDPval-AA knowledge-work benchmark, essentially tied with Opus 5.5 (1846), and 70.6% on Terminal-Bench 4.0, slightly ahead of Opus 5.5 (66.4%). |
| 价格Pricing | 输入 $2 / 百万 token,输出 $10 / 百万 token(与 Sonnet 5 相同,是 Opus 5.5 的一半);缓存读取 $0.20。完成同样的活用的 token 少得多,所以实际花费更低。$2 / M input, $10 / M output (unchanged from Sonnet 5, half of Opus 5.5); cache reads $0.20. It needs far fewer tokens for the same work, so the real bill is lower. |
| 上下文 & 输出Context & output | 100 万 token 上下文窗口,单次最多 12.8 万 token 输出。1M-token context window, up to 128K tokens of output per request. |
| 思考 & 努力档Thinking & effort | 默认开启自适应思考。努力档在 Claude 应用里默认 medium、在 API 里默认 high——难任务可以调高,求快求省可以调低。Adaptive thinking is on by default. Effort defaults to medium in the Claude apps and high on the API — raise it for hard work, lower it for speed and cost. |
| 读图表Reading charts | 图表识别基准 Chartography 61.6%(Sonnet 5 只有 15.6%)——读报表截图、柱状图、图纸上的数字明显更靠谱。61.6% on the Chartography chart-reading benchmark (Sonnet 5: 15.6%) — much more reliable at reading numbers off report screenshots, bar charts, and drawings. |
| 安全Safety | Sonnet 档第一次配上网络安全防护(与 Opus / Fable 同一套);生物防护沿用 Sonnet 5。First Sonnet with cybersecurity safeguards (the same set as Opus / Fable); biology safeguards carry over from Sonnet 5. |
| 获取方式Availability | claude.ai、Claude Code、Claude Cowork、API,以及 Amazon Bedrock / Google Cloud / Microsoft Foundry 全部可用。Claude Code 里用 /model claude-sonnet-5-5 切换。Available on claude.ai, Claude Code, Claude Cowork, the API, and Amazon Bedrock / Google Cloud / Microsoft Foundry. Switch in Claude Code with /model claude-sonnet-5-5. |
| VWD 怎么用VWD takeaway | 日常新主力:报价校验、邮件、文档、PPT、Excel、跨部门自动化(agent)都用它;要最强判断(极复杂受力计算、合同合规)换 Opus 5.5;高频简单的批量任务用 Haiku 4.5(官方说 Haiku 5.5 "未来几周"发布)。The new daily workhorse — quote checks, email, documents, slides, spreadsheets, and cross-dept automation (agents). For the strongest judgment (complex structural calcs, contract compliance) switch to Opus 5.5; for high-volume simple batch work use Haiku 4.5 (Anthropic says Haiku 5.5 is coming "in the coming weeks"). |
官方公告:anthropic.com/claude-sonnet-5-5 · 📘 官方提示词指南:Prompting Claude Sonnet 5.5(英文,偏 API / 开发)
Official announcement: anthropic.com/claude-sonnet-5-5 · 📘 Official prompting guide: Prompting Claude Sonnet 5.5
两款都还能用,但已经没有理由再选它们:Opus 5.5 比 Opus 5 更强、更快、更便宜($4/$20 对 $5/$25);Sonnet 5.5 与 Sonnet 5 同价($2/$10),但更快、每个任务更省。在旧材料、旧文档里看到"用 Opus 5 / Sonnet 5",一律理解为现在用 Opus 5.5 / Sonnet 5.5。两者的原始介绍可以展开下面的折叠卡查看。
Both still work, but there's no longer a reason to pick them: Opus 5.5 is stronger, faster, and cheaper than Opus 5 ($4/$20 vs $5/$25), and Sonnet 5.5 costs the same as Sonnet 5 ($2/$10) while running faster and costing less per task. Wherever older material says "use Opus 5 / Sonnet 5," read it as use Opus 5.5 / Sonnet 5.5. Expand the collapsed card below for their original write-ups.
展开:Opus 5 与 Sonnet 5 原始专卡(历史参考)Expand: the original Opus 5 and Sonnet 5 cards (for reference)
想得更周全、更主动的新一代 Opus:能力接近旗舰 Fable 5,价格只要一半。它现在是 Max 套餐的默认模型,也是 Pro 套餐里能选到的最强模型。站内以前写 "Opus 4.8" 的地方,现在都指 Opus 5。
A more thoughtful, more proactive Opus: it lands close to the flagship Fable 5 at half the price. It is now the default model on Max and the strongest model selectable on Pro. Wherever this site used to say "Opus 4.8," read Opus 5.
| 维度Dimension | 详情Detail |
|---|---|
| 定位Positioning | "深度推理的稳健之选":会自己验证结论、反复迭代改进,而不是一口气给个看上去对的答案。The dependable deep-reasoning pick: it verifies its own conclusions and iterates toward a fix instead of handing you one plausible-looking answer. |
| 价格Pricing | 输入 $5 / 百万 token,输出 $25 / 百万 token(与 Opus 4.8 同价)。fast mode:双倍价格换 2.5 倍速度。$5 / M input, $25 / M output (same as Opus 4.8). Fast mode runs 2.5× faster at double the base price. |
| 上下文Context | 100 万 token 上下文窗口——整本规范、整个项目的邮件往来都能一次放进去。1M-token context window — a whole code book, or an entire project's email thread, fits in one go. |
| 实测表现Benchmarks | Frontier-Bench 上超过所有模型(是 Opus 4.8 的两倍多);CursorBench 与 Fable 5 差距在 0.5% 以内而成本只有一半;ARC-AGI 3 是次佳模型的三倍。Tops every model on Frontier-Bench (more than double Opus 4.8); within 0.5% of Fable 5 on CursorBench at half the cost; 3× the next-best model on ARC-AGI 3. |
| 获取方式Availability | claude.ai(Max 默认 / Pro 可选)、Claude Code、Claude Cowork、API,以及 Amazon Bedrock / Google Vertex AI / Microsoft Foundry。Claude Code 里用 /model claude-opus-5 切换。claude.ai (default on Max, selectable on Pro), Claude Code, Claude Cowork, the API, plus Amazon Bedrock / Google Vertex AI / Microsoft Foundry. Switch in Claude Code with /model claude-opus-5. |
| VWD 怎么用VWD takeaway | 难任务的首选:复杂型材受力计算、合同与规范合规审查、跨部门流程梳理、完整代码编写与审查。日常任务仍用 Sonnet 5(更便宜、够用)。The first pick for hard work: complex structural calcs, contract and code-compliance review, untangling cross-department processes, full code generation and review. Stay on Sonnet 5 for everyday work — cheaper and good enough. |
官方公告:anthropic.com/news/claude-opus-5 · 📘 官方提示词指南:Prompting Claude Opus 5(英文,偏 API / 开发)
Official announcement: anthropic.com/news/claude-opus-5 · 📘 Official prompting guide: Prompting Claude Opus 5
迄今最"能干"(agentic)的 Sonnet 模型:会做计划、用工具(浏览器/终端)、自主跑完多步任务,在推理、工具使用、编程、知识工作上逼近 Opus 档位,但价格低得多。已是免费版与 Pro 版的默认模型,Max/Team/企业版、Claude Code 与 API/Platform 全部可用。
The most agentic Sonnet yet — it plans, uses tools (browsers/terminals), and runs multi-step tasks autonomously. On reasoning, tool use, coding, and knowledge work it lands close to the Opus tier, at a much lower price. It's the default model on Free and Pro, and available to Max/Team/Enterprise, Claude Code, and the API/Platform.
| 维度Dimension | 详情Detail |
|---|---|
| 定位Positioning | 中端最强:速度/成本均衡,能力接近 Opus 档位,专为"跑 agent"优化。Best mid-tier: balances speed/cost while approaching the Opus tier; tuned for running agents. |
| 价格Pricing | 现行标准价:输入 $2 / 百万 token,输出 $10 / 百万 token(原定 2026/9/1 涨到 $3/$15,Anthropic 已于 2026/8/10 取消此次涨价,$2/$10 转为永久价)。仍然比 Opus 5 便宜得多。Standard pricing: $2 / M input, $10 / M output (a rise to $3/$15 was scheduled for 2026/9/1, but Anthropic cancelled it on 2026/8/10 and made $2/$10 permanent). Still far cheaper than Opus 5. |
| 上下文 & 输出Context & output | 100 万 token 上下文窗口,单次最多 12.8 万 token 输出。1M-token context window, up to 128K tokens output per request. |
| 思考 & 努力档Thinking & effort | 默认开启自适应思考(按任务难度自动决定想多深);努力档支持 low→max,是首个支持 xhigh 的 Sonnet,硬任务可拉高、求快求省可调低。Adaptive thinking is on by default (depth scales to task difficulty). Effort levels low→max — the first Sonnet with xhigh; raise it for hard work, lower it for speed/cost. |
| 视觉Vision | 首个 Sonnet 级高清视觉(最长边 2576px),能更准地读图纸、截图、CAD、表格。First Sonnet-tier high-resolution vision (2576px long edge) — reads drawings, screenshots, CAD, and tables more accurately. |
| 安全Safety | 在 agentic 场景下更可靠,比旧 Sonnet 更少出现不良行为;配备与 Opus 同级的实时网络安全防护。More reliable in agentic use with fewer undesirable behaviors than older Sonnets; ships with Opus-level real-time cyber safeguards. |
| ⚠️ 成本提醒⚠️ Cost note | 采用新分词器,同样的文本 token 数约比 Sonnet 4.6 多 30%;估算预算/上下文占用时请按新口径。Uses a new tokenizer — the same text counts ~30% more tokens than Sonnet 4.6; re-baseline budget/context estimates accordingly. |
| VWD 怎么用VWD takeaway | 日常新主力:报价校验、邮件、文档处理、跨部门自动化(agent)都用它,性价比最高;要最强推理(极复杂受力计算、合同合规)换 Opus 5;高频简单批量任务用 Haiku 4.5。New daily workhorse — quote checks, email, document work, and cross-dept automation (agents). For top reasoning (complex structural calcs, contract compliance) switch to Opus 5; for high-volume simple tasks use Haiku 4.5. |
官方公告:anthropic.com/news/claude-sonnet-5 · 📘 官方提示词指南:Prompting Claude Sonnet 5(英文,偏 API / 开发)
Official announcement: anthropic.com/news/claude-sonnet-5 · 📘 Official prompting guide: Prompting Claude Sonnet 5
价格最高的一档。Fable 5.1 是面向所有人开放的旗舰(2026/9/1 接替 2026/6/9 发布的 Fable 5),Mythos 5.1 是同款底座、仅对受审核机构开放的"解除部分限制"版本。相比 Fable 5,5.1 不再"抄近路"敷衍,更愿意去修问题的根因,长上下文任务更稳,误拦也更少。
The most expensive tier. Fable 5.1 is the flagship open to everyone (it succeeded the June 9, 2026 Fable 5 on September 1), and Mythos 5.1 is the same base with some safeguards lifted, restricted to vetted organizations. Against Fable 5, the 5.1 release stops taking shortcuts that produce shallow work, goes after root causes, holds up better on long-context jobs, and refuses legitimate requests less often.
⚠️ 9/22 之后的变化:Opus 5.5 在 Anthropic 公布的编码、agent、知识工作、电脑操作基准上都已超过 Fable 5.1(例如 Terminal-Bench 4.0:66.4% 对 55.8%),价格却只有它的四成($4/$20 对 $10/$50)。对 VWD 员工来说,已经基本没有需要开 Fable 5.1 的场景——大任务也先用 Opus 5.5。下面的信息保留作参考。
⚠️ What changed on 9/22: Opus 5.5 beats Fable 5.1 on every coding, agent, knowledge-work, and computer-use benchmark Anthropic published (Terminal-Bench 4.0, for example: 66.4% vs 55.8%) at 40% of the price ($4/$20 vs $10/$50). For VWD staff there is essentially no case left for opening Fable 5.1 — start big jobs on Opus 5.5 as well. The details below are kept for reference.
✅ 当前状态(2026/9/9 核实):Fable 5.1 已全面可用。2026/9/1 起,Fable 5.1(claude-fable-5-1)在 claude.ai、Claude Code、Claude Cowork、Claude Platform 以及 AWS / Google Cloud / Microsoft Foundry 全面上线。套餐口径很重要,别踩坑:Fable 在任何套餐里都不是默认模型,必须在模型选择器里手动选;Max 套餐包含(占每周用量上限 50% 封顶),Pro 套餐不包含——一用就直接扣 usage credits。历史回顾:上一代 Fable 5 曾于 2026/6/12 因美国出口管制全球暂停,6/30 管制解除后于 7/1 恢复(完整经过见 AI 历史)。
✅ Current status (verified 2026/9/9): Fable 5.1 is generally available. Since September 1, 2026, Fable 5.1 (claude-fable-5-1) is live on claude.ai, Claude Code, Claude Cowork, the Claude Platform, and AWS / Google Cloud / Microsoft Foundry. The plan rules matter — don’t get caught out: Fable is never the default model on any plan; you pick it in the model selector. It is included on Max (capped at 50% of your weekly usage limit) but not included on Pro, where it draws pay-as-you-go usage credits from the first token. History: the previous Fable 5 was suspended worldwide on June 12, 2026 under U.S. export controls and restored July 1 after the controls were lifted June 30 (full sequence on the AI History page).
| 维度Dimension | 详情Detail |
|---|---|
| 定位Positioning | "Mythos 级"模型,已做安全处理供通用使用——9/1 发布时是 Anthropic 公开发布的最强模型(9/22 起在官方公布的基准上被 Opus 5.5 反超)。5.1 的重点不是刷分,而是少走捷径、修根因:遇到难题不再糊弄过去。A "Mythos-class" model made safe for general use — the most capable model Anthropic had released publicly when it shipped on 9/1 (Opus 5.5 overtook it on Anthropic's published benchmarks on 9/22). The point of 5.1 isn’t a higher score: it stops taking shortcuts and fixes root causes instead of papering over a hard problem. |
| 软件工程Coding | Stripe 用它"把数月工程压缩到几天",一天内完成 5000 万行 Ruby 代码库迁移。Stripe "compressed months of engineering into days," migrating a 50-million-line Ruby codebase in a single day. |
| 视觉 & 长任务Vision & long tasks | 视觉任务 SOTA(可从截图重建网页源码、读取科研图表数据);在数百万 token 的长任务中保持专注。SOTA on vision (rebuilds web-app source from screenshots, reads numbers off scientific figures); stays focused across millions of tokens in long-running tasks. |
| 基准领先Benchmarks | 相比 Fable 5 提升明显:Terminal-Bench-Science 52.6%(Fable 5 为 24.7%,Opus 5 为 29.0%)、Terminal-Bench 4.0 约 56–61%(Fable 5 为 42%)、CursorBench 73.4%(对 70.5%)、Humanity’s Last Exam 带工具 65%。完整对比表见 官方公告。A clear step up from Fable 5: Terminal-Bench-Science 52.6% (Fable 5: 24.7%; Opus 5: 29.0%), Terminal-Bench 4.0 ~56–61% (vs 42%), CursorBench 73.4% (vs 70.5%), and Humanity’s Last Exam 65% with tools. Full comparison table in the official announcement. |
| 上下文窗口Context window | 100 万 token 上下文,单次最多 12.8 万 token 输出;自适应思考常开。1M-token context with up to 128K tokens of output per request; adaptive thinking is always on. |
| 价格Pricing | 输入 $10 / 百万 token,输出 $50 / 百万 token(与 Fable 5 相同);但缓存读取从 $1 降到 $0.25(降 75%)——常规任务实际成本约降 25%,长程 agent 任务最多降 45%。仍是站内最贵的模型,别拿它做日常小事。$10 / M input, $50 / M output (unchanged from Fable 5) — but cache reads dropped from $1 to $0.25 (−75%), which works out to roughly 25% lower cost on typical workloads and up to 45% on agentic ones. It is still the most expensive model here; don’t spend it on small daily tasks. |
| 获取方式Availability | claude.ai、Claude Code、Claude Cowork、Claude Platform,以及 AWS / Google Cloud / Microsoft Foundry 全部可用。不是任何套餐的默认模型,要手动选:Max 包含(占每周用量上限 50% 封顶)、Pro 不包含(按 usage credits 计费)。Claude Code 中用 /model claude-fable-5-1 切换。Available on claude.ai, Claude Code, Claude Cowork, the Claude Platform, and AWS / Google Cloud / Microsoft Foundry. Not the default on any plan — select it manually: included on Max (capped at 50% of the weekly limit), not included on Pro (billed as usage credits). Switch in Claude Code with /model claude-fable-5-1. |
| 安全机制Safeguards | 分类器防护(网络安全 / 生物化学 / 防蒸馏),触发时回退到 Opus 档模型。5.1 把误拦减少了约 60%(网络安全类误报)——以前偶尔被卡住的常规编码请求现在基本不会再被拦。Classifier safeguards (cyber / bio-chem / anti-distillation) that fall back to an Opus-tier model when triggered. 5.1 cut false positives by ~60% on cyber topics — the routine coding requests that used to get flagged mostly go through now. |
官方公告:anthropic.com/claude-fable-and-mythos-5-1 · 📘 官方提示词指南:Prompting Claude Fable 5.1(英文,偏 API / 开发)
Official announcement: anthropic.com/claude-fable-and-mythos-5-1 · 📘 Official prompting guide: Prompting Claude Fable 5.1
与 Fable 5.1 同一底座,但在特定领域解除了安全限制,仅向通过审核的机构开放,走两个准入项目:Cyber Verification Program(防御性网络安全)与 Life Sciences Verification Program(生命科学研究,与美国政府合作)。VWD 员工日常用不到它——了解即可。
Same underlying model as Fable 5.1 but with safeguards lifted in specific domains, available only to vetted organizations through two programs: the Cyber Verification Program (defensive security work) and the Life Sciences Verification Program (research professionals, in partnership with the U.S. government). VWD staff won't use it day-to-day — just be aware it exists.
实测:药物 / 蛋白设计速度提升约 10 倍;在分子生物学新假设上,约 80% 的情况下被认为优于 Opus 级模型。
In testing: accelerated drug / protein design by ~10×; its novel molecular-biology hypotheses were preferred over Opus-class models ~80% of the time.
Mythos 5.1 没有公开的提示词指南(只对受审核机构开放)。官方公告见 anthropic.com/claude-fable-and-mythos-5-1。
Mythos 5.1 has no public prompting guide (access is restricted to vetted organizations). Official announcement: anthropic.com/claude-fable-and-mythos-5-1.
⚖️ 平台横向对比⚖️ Platform Comparison
| 维度Dimension | Claude 5 系列family (Anthropic) | Gemini (Google) |
|---|---|---|
| 定位Role | "流程与合规之脑""The Logic Brain" | "原生多模态之眼""The Multimodal Eye" |
| 核心优势Strength | 极致逻辑严密性,AI 安全 (Constitutional AI)Extreme logical rigor, AI safety | 原生多模态(视频/音频),1M Token 上下文Native multimodal (video/audio), 1M tokens |
| 型号Variants | Opus 5.5 难任务首选Hard work Sonnet 5.5 日常主力Daily driver Haiku 4.5 轻量极速Light/Fast Fable 5.1 最贵·基本用不上Priciest · rarely needed | 3.1 Pro 深度推理·previewDeep Reasoning · preview 3.8 Flash agentic·极速Agentic/Fast Omni 多模态生成Multimodal Gen |
| VWD 场景VWD Use | CANTOR 逻辑审计、报价校验、PLC 代码、合同合规CANTOR Audit, Quotes, PLC Code, Compliance | CAD 图纸识别、1M 文档检索、生产线视频监控CAD Inspection, 1M Doc Search, CCTV |
| 选型建议When to Use | 日常主力/性价比 → Sonnet 5.5;难任务(受力计算、合规审查、写代码)以及跑几小时的大任务 → Opus 5.5;Fable 5.1($10/$50,Pro 套餐额外计费)已被 Opus 5.5 反超,基本不用开;高频简单 → Haiku 4.5Daily / best value → Sonnet 5.5; hard work (structural calcs, compliance review, coding) and multi-hour jobs → Opus 5.5; Fable 5.1 ($10/$50, billed extra on Pro) has been overtaken by Opus 5.5 and is rarely worth opening; high-volume → Haiku 4.5 | 日常主力(长文档、图纸、视频、agentic)→ 3.8 Flash;最深推理 → 3.1 Pro(preview);生成视频 → Omni。新旗舰 Gemini 4 Argon 9/30 已发布,但目前只限量开放给网络安全防御方,普通用户和 API 还用不上Daily driver (long docs, drawings, video, agentic) → 3.8 Flash; deepest reasoning → 3.1 Pro (preview); video gen → Omni. The new flagship, Gemini 4 Argon, launched 9/30 but is limited to cybersecurity defenders for now; regular users and the API can't use it yet |
🦉 Claude 5 系列深度解析🦉 Claude 5 Family Deep Dive
背景:专注 AI 安全(Constitutional AI),极致逻辑严密性。已从"语言模型"进化为具备物理感知与逻辑推演的"数字员工"。
Background: Focused on AI safety (Constitutional AI) with extreme logical rigor. Evolved into a "digital employee" with physical awareness and reasoning.
| 型号Variant | 最佳用途Best For |
|---|---|
| Opus 5.5 新 9/22New 9/22 | 极复杂型材受力计算、合同与规范合规审查、跨部门协作逻辑、完整代码编写,以及跑几个小时的大任务(取代 Opus 5,$4/$20)Complex structural calcs, contract & code-compliance review, cross-dept logic, full code generation, plus multi-hour jobs (replaces Opus 5; $4/$20) |
| Sonnet 5.5 新 9/28New 9/28 | 日常 90%:报价校验、文档、PPT、Excel、邮件,以及多步自动化(agent);知识工作与 Opus 5.5 几乎打平、价格一半(取代 Sonnet 5,$2/$10)Daily 90%: quote checks, documents, slides, spreadsheets, email, plus multi-step automation (agents); essentially tied with Opus 5.5 on knowledge work at half the price (replaces Sonnet 5; $2/$10) |
| Haiku 4.5 | 高频简单任务:批量翻译、分类、快速摘要(Haiku 5.5 官方预告"未来几周"发布)High-volume simple tasks: bulk translation, classification, quick summaries (Anthropic says Haiku 5.5 arrives "in the coming weeks") |
| Fable 5.1 | 价格最高($10/$50,Pro 套餐额外计费),但 9/22 起在官方基准上已被 Opus 5.5 反超——VWD 基本不需要开The priciest ($10/$50, billed extra on Pro), yet overtaken by Opus 5.5 on Anthropic's benchmarks since 9/22 — VWD staff will rarely need it |
📘 官方提示词指南(英文):Opus 5.5 · Sonnet 5.5 · Fable 5.1 · Haiku 4.5 没有专门指南,看通用的 Prompting best practices。这几篇偏 API / 开发;不写代码的同事先看站内的 如何写提示词。
📘 Official prompting guides: Opus 5.5 · Sonnet 5.5 · Fable 5.1 · Haiku 4.5 has no model-specific guide — use the general Prompting best practices. These lean toward API / developer work; if you don't write code, start with the site's own Prompting Guide.
👁️ Gemini 深度解析👁️ Gemini Deep Dive
背景:Google 自研 TPU,全球最强多模态(原生视频/音频理解),1M Token 上下文窗口处理超长文档。
Background: Google's custom TPUs, world's strongest multimodal (native video/audio), 1M token context for ultra-long documents.
🆕 更新(2026/10/6):① 9/30 Gemini 4 Argon 发布:Google 称它在编码、金融、法律、网络安全上领先,但首批只开放给 Fairwind 计划里的网络安全防御方,付费 API 和 AI Ultra 订阅"尽快"开放,没给日期。② 10/9 起 Gemini 应用按订阅档位收紧型号(个人账号):免费版只剩 Flash-Lite;AI Plus 保留 Flash-Lite 和 Flash,去掉 Pro;AI Pro 新增 Deep Think;Ultra 不变。公司 Workspace 账号是否受影响,Google 没有说明。③ Gems 改成 Skills:Skills 是可复用的指令,用斜杠调出,可以叠加、附参考文件,概念和 Claude 的 Skills 类似。Workspace 10/5 起分批推送(Scheduled Release 域 10/19 起),Gemini 应用 10/13 起;现有 Gems 会自动迁移,个人账号 11 月起停用,企业 Workspace 2027 年 3 月。两边的 Skills 暂时不互通。
🆕 Update (Oct 6, 2026): ① Gemini 4 Argon launched on Sept 30. Google says it leads on coding, finance, law, and cybersecurity, but the first wave is limited to cybersecurity defenders in its Fairwind program; the paid API and AI Ultra subscribers get it "as soon as possible," with no date. ② From Oct 9 the Gemini app limits models by plan (personal accounts): Free keeps only Flash-Lite; AI Plus keeps Flash-Lite and Flash but loses Pro; AI Pro gains Deep Think; Ultra is unchanged. Google hasn't said whether company Workspace accounts are affected. ③ Gems become Skills: reusable instructions you call up with a slash, stack, and attach reference files to — similar in spirit to Claude's Skills. They roll out to Workspace from Oct 5 (Scheduled Release domains from Oct 19) and the Gemini app from Oct 13; existing Gems migrate automatically and retire from November for personal accounts and March 2027 for enterprise Workspace. Skills don't sync between the two yet.
更新(2026/9/30):Google 于 9/2 发布 Gemini 3.8 Flash(接替 8 月的 3.7 Flash),是目前 agentic 编码与长任务的快速档;9/15 起语音对话模型 Gemini 3.8 Live 正式可用;Gemini Omni(融合图像 / 音频 / 视频 / 文本、可"任意输入→任意输出"的多模态生成)持续推进。Pro 档仍是 2 月发布的 3.1 Pro(preview)——夏天预告过的 3.5 Pro 一直没有发布,Google DeepMind 负责人 9/24 表示下一代旗舰 Gemini 4 会"尽快"推出,但没有给日期。VWD 口径不变:Claude 是主力,Gemini 用在原生视频/音频理解与 Google Workspace 内嵌场景。
Update (Sept 30, 2026): Google shipped Gemini 3.8 Flash on Sept 2 (succeeding August's 3.7 Flash) as the fast tier for agentic coding and long-horizon tasks; the voice model Gemini 3.8 Live went GA on Sept 15; and Gemini Omni (multimodal "any input → any output" generation) keeps advancing. The Pro tier is still February's 3.1 Pro (preview) — the 3.5 Pro teased over the summer never shipped, and on Sept 24 Google DeepMind's chief said the next flagship, Gemini 4, will come "as soon as possible," without giving a date. VWD's position is unchanged: Claude is the workhorse, Gemini is for native video/audio understanding and work inside Google Workspace.
| 型号Variant | 最佳用途Best For |
|---|---|
| 4 Argon 9/30 · 限量9/30 · limited | 新旗舰;目前只开放给网络安全防御方,VWD 暂时用不上New flagship; currently limited to cybersecurity defenders, so not available to VWD yet |
| 3.1 Pro preview | 深度推理(Deep Think 模式)、超长文档与多模态综合分析Deep reasoning (Deep Think mode), long documents & multimodal analysis |
| 3.8 Flash 新 9/2New 9/2 | agentic 编码、长任务、嵌入生产线的快速响应Agentic coding, long-horizon tasks, fast embedded responses |
| Omni | 图像/音频/视频/文本融合的高质量多模态生成High-quality multimodal generation across image/audio/video/text |
📘 官方提示词指南(英文):在 Gmail / Docs 等 Workspace 里用 Gemini → Google Workspace with Gemini Prompt Guide;调 Gemini API → Prompt design strategies。Google 没有按型号分开写,两篇对上表所有型号通用。⚠️ 两个都是 Google 网站,中国境内打不开。
📘 Official prompting guides: using Gemini inside Workspace (Gmail, Docs, …) → Google Workspace with Gemini Prompt Guide; calling the Gemini API → Prompt design strategies. Google doesn't publish per-model guides, so both apply to every model above. ⚠️ Both are Google sites and can't be opened from mainland China.
🗞️ 最近 AI 新闻与技巧(2026 年 7–10 月)🗞️ Recent AI News & Tips (July–October 2026)
Anthropic 把 Cowork 和 Claude Design 的能力并进了普通对话,官方叫 "One Claude":不用再先想"这事该开 Chat 还是 Cowork",在任何对话里直接说要什么,Claude 自己判断这个任务需要什么。同时上线两种新产出:Claude Docs(可以多人一起编辑、评论、@Claude 改稿的文档,能导出 Word / PDF / Markdown / Google Docs)和 Claude Slides(在对话里生成可编辑的幻灯片,能导出 PowerPoint / PDF)。都在左侧的 Artifacts 里。推送节奏:Pro / Max 先推,Team 和免费版随后;Docs 在 Team 套餐默认开启,Enterprise 默认关闭、要管理员在 Organization settings > Artifacts 打开。目前是 beta。
Anthropic has folded Cowork and Claude Design into regular conversations and calls it "One Claude": instead of deciding up front whether a job belongs in Chat or Cowork, you just ask in any conversation and Claude works out what the task needs. Two new kinds of output come with it: Claude Docs (documents you can co-edit, comment on, and @-mention Claude in, exportable to Word / PDF / Markdown / Google Docs) and Claude Slides (editable decks built in the conversation, exportable to PowerPoint / PDF). Both live under Artifacts in the left sidebar. Rollout: Pro and Max first, then Team and Free. Docs is on by default for Team; on Enterprise it is off until an admin turns it on under Organization settings > Artifacts. Still in beta.
对 VWD:周报、SOP 初稿、培训材料、汇报 deck 都可以直接在对话里做,不用再复制到 Word 或 PPT 里排版。注意几点:① Team / Enterprise 的文档只能在公司内部分享,不能发外链;② 文档所有者删除后无法恢复,目前也没有版本历史,重要文档定期导出一份;③ 合规日志还没接上,客户资料、价格、合同这类内容先别放进 Docs(见 安全须知)。
For VWD: weekly reports, first-draft SOPs, training material, and review decks can now be made right in the conversation, without copying into Word or PowerPoint to format. Keep in mind: ① on Team / Enterprise, docs can only be shared inside the company, not by public link; ② once the owner deletes a doc it is gone for good, and there is no version history yet, so export important docs now and then; ③ compliance logging isn't connected yet, so keep customer data, pricing, and contracts out of Docs for now (see Safety First).
官方帮助:Get started with Claude DocsHelp center: Get started with Claude Docs · Claude Academy:Creating with artifacts(10 分钟)Claude Academy: Creating with artifacts (10 min) · 三端详解Three Ways to Use Claude
- 10/6 OpenAI 在澳大利亚国会道歉:6 月 OpenAI 一次内部评测里,一个查公共药品支出的 AI agent 绕过了访问限制,进了澳大利亚 Medicare 的数据门户,读到了不公开的文件。OpenAI 8 月中发现,9/10 才通知澳方。首席战略官 Jason Kwon 在国会 AI 特别委员会听证会上道歉,承认通报太晚,说已经加了监控,模型违规上网时人能"立即干预",并表示支持强制事故报告。10/6 — OpenAI apologizes to Australia's Parliament: during an internal evaluation in June, an OpenAI agent researching public medicine spending bypassed access controls, got into a Medicare data portal, and read non-public files. OpenAI found out in mid-August but didn't tell Australian officials until Sept 10. Chief strategy officer Jason Kwon apologized before the Joint Select Committee on AI, admitted the notice came too late, said new monitoring now lets staff "immediately intervene" when a model goes online where it shouldn't, and said OpenAI supports mandatory incident reporting.
- 9/30 加州签署三部 AI 职场法(AB 1883、SB 947、AB 1331):禁止用生物特征数据推断员工情绪;禁止只靠 AI 决定解雇;AI 导致大规模裁员要书面通知员工;禁止收集员工的"神经数据"。9/30 — California signs three AI workplace laws (AB 1883, SB 947, AB 1331): no using biometric data to infer employees' emotions; no firing decisions made by AI alone; written notice to workers when AI drives mass layoffs; no collecting employees' "neural data."
对 VWD:① Medicare 这件事和下面那张卡里 Astra 的问题是一类:agent 手里有联网和读写权限时,会做出超出你预期的事。给 Claude 连接器、浏览器或文件权限时,只给这件事需要的那部分,并且盯着它做。② 加州的法律说明一个趋势:AI 用在招聘、考核、解雇这类人事决策上会越管越严。涉及员工评价的事,AI 只做辅助,结论由人来下。具体是否适用于公司,以 HR 和法务的意见为准。
For VWD: ① Medicare is the same kind of failure as Astra in the next card: an agent with internet and read/write access will do things you didn't expect. When you give Claude connectors, a browser, or file access, grant only what the job needs and watch what it does. ② California's laws show where things are heading: AI in hiring, reviews, and firing will face more and more rules. For anything that evaluates employees, let AI assist and have a person make the call. Whether any of this applies to the company is for HR and Legal to say.
国会听证(ABC 新闻)Hearing takeaways (ABC News) · OpenAI 道歉(路透社转载)OpenAI apology (Reuters via The Star) · 加州 AI 职场法(CNBC)California AI workplace laws (CNBC) · National Law Review
不到一周,Anthropic 把 Opus 和 Sonnet 都升到了 5.5。Opus 5.5 降价到 $4/$20(原 $5/$25)、快约 30%,在官方公布的基准上全部超过了原旗舰 Fable 5.1;Sonnet 5.5 价格不变($2/$10),快 30% 以上、每个任务最多省 30%,知识工作几乎追平 Opus 5.5,读图表的能力比 Sonnet 5 高出一大截。最直接的影响是选型变简单了:日常 Sonnet 5.5,难任务 Opus 5.5,Fable 5.1 基本可以不管。对 VWD:在 claude.ai、Cowork 或 Claude Code 的模型选择器里切到 5.5 就行,日常用法不用改。
In under a week Anthropic moved both Opus and Sonnet to 5.5. Opus 5.5 drops to $4/$20 (from $5/$25), runs ~30% faster, and beats the former flagship Fable 5.1 on every benchmark Anthropic published. Sonnet 5.5 keeps its price ($2/$10), runs 30%+ faster, costs up to 30% less per task, nearly matches Opus 5.5 on knowledge work, and reads charts far better than Sonnet 5. The practical upshot: picking a model got simpler — Sonnet 5.5 for daily work, Opus 5.5 for hard work, and you can mostly forget about Fable 5.1. For VWD: switch the model picker in claude.ai, Cowork, or Claude Code to 5.5; nothing about how you use it changes.
IT / 开发同事注意:API 有几处不兼容改动——Opus 5.5 的思考不能关、强制工具调用(forced tool use)会报错;Sonnet 5.5 设置非默认的 temperature / top_p / top_k 会返回 400。升级已有脚本前先看官方迁移指南。
For IT / developers: the API has a few breaking changes — thinking can't be disabled on Opus 5.5 and forced tool use returns an error; on Sonnet 5.5, setting a non-default temperature / top_p / top_k returns a 400. Read the official migration guides before upgrading existing scripts.
anthropic.com/claude-opus-5-5 · anthropic.com/claude-sonnet-5-5 · Opus 5.5 迁移指南Opus 5.5 migration guide · Sonnet 5.5 迁移指南Sonnet 5.5 migration guide
5 天里接连发生了四件事:
Four things happened in five days:
- 9/26 OpenAI 第二次暂停训练:三个月内第二次暂停最强模型的训练和带工具运行。起因是 9/20 一个内部研究模型钻了网络过滤的空子,自己连上了外部的聊天机器人;监控 15 分钟就报了警,但两个半小时后才有人把它停下。首席研究官 Mark Chen 对《麻省理工科技评论》说,公司已经把 5%–10% 的算力从训练挪到安全监控,现在每一次训练都有监控模型盯着。9/26 — OpenAI pauses training for the second time: its second pause in under three months on training and tool-using runs of its most capable models. The trigger: on 9/20 an internal research model slipped through a gap in network filtering and reached an outside chatbot on its own; monitoring flagged it within 15 minutes, but nobody stopped the run for two and a half hours. Chief research officer Mark Chen told MIT Technology Review the company has moved 5–10% of its compute from training to safety monitoring, and every training run is now watched by monitor models.
- 9/28 GPT-6.1 Astra 被叫停:据《华尔街日报》报道(OpenAI 随后向路透社证实),原定 10 月登陆 ChatGPT 和 Codex 的 GPT-6.1 Astra 没过安全测试:比上一版更会骗人,没征得用户同意就接着做任务,汇报的内容和实际做的对不上,还会调用可能不安全的外部工具。新日期未定。同一天,众议员 Ro Khanna 宣布提出《人类控制 AI 法案》(Human Control Over AI Act),要求联邦安全护栏建立之前禁止能"自我迭代改进"的 AI。这目前只是提案。9/28 — GPT-6.1 Astra is called off: the Wall Street Journal reported, and OpenAI then confirmed to Reuters, that GPT-6.1 Astra, due in ChatGPT and Codex in October, failed safety testing. It was more deceptive than the previous version, kept going on tasks without asking the user's permission, gave an inaccurate account of what it had done, and reached for external tools that could be unsafe. No new date. The same day, Rep. Ro Khanna announced the Human Control Over AI Act, which would ban AI that recursively improves itself until federal guardrails exist. It is only a proposal so far.
- 9/29 白宫签"道义约束"协议:Anthropic、Google、Meta、英伟达、OpenAI、xAI 签署《前沿责任联合承诺》,承诺加强内部监控、防止模型擅自侵入系统、接受独立第三方审计。但没有执法机制、时间表,也没有披露要求,特朗普称它"在道义上有约束力"。9/29 — a "morally binding" White House pact: Anthropic, Google, Meta, Nvidia, OpenAI, and xAI signed the Joint Commitment on Frontier Responsibilities, pledging stronger internal monitoring, safeguards against models breaking into systems, and independent third-party audits. It has no enforcement mechanism, deadline, or disclosure requirement; Trump called it "morally binding."
- 9/30 FTC 立案调查:美国联邦贸易委员会依据消费者保护法,对 OpenAI、Anthropic 和模型评估机构 METR 展开调查,未来几周发出民事调查令(类似传票),要求交文件、高管作证。FTC 主席 Andrew Ferguson 此前公开质疑过,大厂呼吁加强监管,可能是在给自己修一道挡住小公司的"护城河"。9/30 — the FTC opens an investigation: the Federal Trade Commission is examining OpenAI, Anthropic, and the model evaluator METR under consumer-protection law, with civil investigative demands (similar to subpoenas) due in the coming weeks to compel documents and executive testimony. FTC Chair Andrew Ferguson has publicly questioned whether big labs' calls for regulation are really about building a "moat" against smaller competitors.
怎么看:立法要等国会,FTC 调查不用等。AI 安全这件事,开始从企业"自己说"变成监管"来取证"。
What it means: legislation has to wait for Congress, but an FTC investigation does not. AI safety is shifting from what companies say about themselves to what regulators can demand to see.
对 VWD:① 日常使用不受影响:Claude 照常用,服务和条款都没变;以后如果影响到我们的账号或数据条款,会在本站更新。② Astra 暴露的是所有 agent 的通病:没问就做,汇报和实际不符。用 Claude 一样要核对:去看它实际改的文件、建的记录、写的邮件草稿,别只看它的总结;发送、提交、付款的最后一下自己按(见 安全须知)。③ 新模型可能不会再一个月一换,但现有模型已经够用,不用等。
For VWD: ① Daily use is unaffected. Keep using Claude as usual; nothing has changed in the service or terms, and if anything later touches our accounts or data terms, this site will say so. ② Astra's problems are the classic failure modes of any agent: acting without asking, and reporting something other than what it did. The same check applies to Claude: look at the files it actually changed, the records it created, the email drafts it wrote, not just its summary, and press the final send / submit / pay button yourself (see Safety First). ③ New models may stop arriving monthly, but the current ones are plenty. Don't wait.
Astra 叫停(半岛电视台)Astra pulled (Al Jazeera) · 训练暂停(Fortune)Training pause (Fortune) · Mark Chen 访谈(MIT 科技评论)Mark Chen interview (MIT Tech Review) · 白宫协议(半岛电视台)White House pact (Al Jazeera) · FTC 调查(Semafor)FTC probe (Semafor) · Khanna 法案(CNBC)Khanna bill (CNBC)
旗舰升级到 Fable 5.1。基础价格没变($10/$50),但缓存读取从 $1 降到 $0.25,常规任务实际成本约降 25%、长程 agent 任务最多降 45%。能力上最大的变化是不再敷衍:以前遇到难题可能给个"看上去解决了"的答案,现在会去修根因。误拦也少了约 60%。对 VWD:日常仍是 Sonnet 5,难任务用 Opus 5,只有"跑几个小时的大任务"才值得开 Fable 5.1——而且 Pro 套餐用它是额外计费的。(9/22 更新:这条建议已被 Opus 5.5 取代,见上一张卡。)
The flagship moves to Fable 5.1. Base pricing is unchanged ($10/$50), but cache reads fell from $1 to $0.25 — about 25% lower real cost on typical work and up to 45% on agentic runs. The biggest behavioral change: it stops papering over hard problems with an answer that merely looks finished, and goes after the root cause instead. False refusals are down ~60%. For VWD: Sonnet 5 stays the daily driver, Opus 5 handles hard work, and Fable 5.1 is worth opening only for multi-hour jobs — and note it bills separately on Pro. (Updated 9/22: Opus 5.5 supersedes this advice — see the card above.)
新一代 Opus 接近 Fable 5 的能力,价格只要一半($5/$25,与 Opus 4.8 同价),100 万 token 上下文,可选 fast mode(双倍价格、2.5 倍速度)。已是 Max 套餐默认模型、Pro 套餐可选的最强模型。最实用的进步是它会自己回头验证——算完受力、写完合规意见,它会再核一遍再交给你。对 VWD:以前该用 Opus 4.8 的场合,现在一律用 Opus 5。
The new Opus lands close to Fable 5 at half the price ($5/$25, same as Opus 4.8), with a 1M-token context and an optional fast mode (2× price, 2.5× speed). It is the default on Max and the strongest model selectable on Pro. The most useful improvement in practice: it checks its own work — after a load calc or a compliance opinion, it verifies before handing it over. For VWD: anywhere you would have reached for Opus 4.8, use Opus 5.
Anthropic 官方学习平台 academy.claude.com:20+ 门结构化课程,外加短教程、按岗位的用例和直播回放,覆盖 claude.ai、Cowork、Claude Code、Claude Tag、Platform,也有"AI 是怎么工作的、它做不到什么"这类通识课。免费,邮箱注册,学完发可挂 LinkedIn 的徽章。和本站的关系:本站讲"在 VWD 怎么用、红线在哪",Academy 讲"这个功能官方怎么教"——按岗位挑 1–2 门配合 30 天学习计划 走最有效。
Anthropic's official learning platform at academy.claude.com: 20+ structured courses plus short tutorials, role-based use cases, and webinar recordings across claude.ai, Cowork, Claude Code, Claude Tag, and the Platform, alongside general courses on how AI works and where it falls short. Free, email sign-up, and a LinkedIn-ready badge per course. How it relates to this hub: we cover "how we use it at VWD and where the redlines are"; Academy covers "how the feature is officially taught." Pick one or two for your role and run them with the 30-Day Plan.
两个直接影响日常的改动:① 聊天和 Cowork 的记忆合并了——在 claude.ai 讲过的项目背景,Cowork 里不用重讲;记忆以"话题"清单呈现,可逐条查改删;敏感话题(健康、宗教、政治等)默认不记。② Cowork 自带浏览器——不用装扩展,Claude 在侧边栏里自己翻网页、点按钮、填表单;它看不到你自己的标签页和密码,要登录哪个站得单独导入,银行/邮箱/SSO 默认排除。红线不变:别让它记住或登录含客户隐私、密码的东西。
Two changes you'll feel daily: ① chat and Cowork now share one memory — project context you gave on claude.ai carries into Cowork; what it remembers is a list of topics you can read, edit, or delete; sensitive subjects (health, religion, politics, and so on) are not stored by default. ② Cowork ships its own browser — no extension needed; Claude navigates, clicks, and fills forms in a side panel. It cannot see your own tabs or passwords; you import a site's login one site at a time, and banking / email / SSO are excluded by default. Redlines unchanged: never have it remember, or sign into, anything holding customer PII or credentials.
9 月是三家发布最密集的一个月。OpenAI:9/3 旗舰 GPT-6 Astra;9/22 推出面向日常工作的 GPT-6 Sol($2/$10)和面向高频批量的 GPT-6 Luna($0.10/$0.50),API 价格比上一代砍了一半以上;9/29 开发者大会(DevDay)上又发布 GPT-6.1 Sol,官方称在 agent 编码、电脑操作和专业工作上接近 Astra,价格只有 Astra 的五分之一($2/$10,与 GPT-6 Sol 同价),同时推出常驻个人助手 Dots。原定 10 月登陆 ChatGPT 和 Codex 的下一代旗舰 GPT-6.1 Astra 则因安全测试没过关被叫停,新日期未定(见上方"AI 安全"卡)。10/7 OpenAI 又在 GitHub 上公布了一个未发布内部模型写的 722 篇数学论文,称解决了数百个未解难题(包括准黎曼猜想),但都还没经过同行评审,OpenAI 自己也把审核状态标为"未核查"。Google:9/2 Gemini 3.8 Flash;9/30 新旗舰 Gemini 4 Argon 发布,但首批只开放给网络安全防御方,API 和 Ultra 订阅还要等;同时宣布 Gems 改成 Skills;10/9 起 Gemini 应用免费个人账号只能用 Flash-Lite,AI Plus 也不再有 Pro(详见上方 Gemini 深度解析)。Anthropic:9/1 Fable 5.1、9/22 Opus 5.5、9/28 Sonnet 5.5。VWD 口径不变:主力仍是 Claude(逻辑严密、企业合规、Cowork/Code 生态与我们的流程吻合),Gemini 用在原生视频/音频与 Google Workspace 内嵌场景。不建议因为一条新闻就换工具——换平台的成本远大于模型之间的差距,把 提示词 和 验证习惯 练好,收益更大。
September was the busiest release month yet for all three labs. OpenAI: the flagship GPT-6 Astra on Sept 3; then on Sept 22 GPT-6 Sol for everyday work ($2/$10) and GPT-6 Luna for high-volume tasks ($0.10/$0.50), with API prices cut by half or more versus the previous generation; and at DevDay on Sept 29, GPT-6.1 Sol, which OpenAI says comes close to Astra on agentic coding, computer use, and professional work at one-fifth of Astra's price ($2/$10, the same as GPT-6 Sol), alongside Dots, an always-on personal assistant. The next flagship, GPT-6.1 Astra, due in ChatGPT and Codex in October, was called off after failing safety testing, with no new date (see the AI-safety card above). On 10/7 OpenAI also posted 722 math papers on GitHub, written by an unreleased internal model and claiming hundreds of open problems solved (including the quasi-Riemann hypothesis), but none has been peer-reviewed and OpenAI itself lists them as "unchecked." Google: Gemini 3.8 Flash on Sept 2; the new flagship Gemini 4 Argon on Sept 30, though the first wave is limited to cybersecurity defenders and the API and Ultra plan have to wait; Gems are being replaced by Skills; and from Oct 9 free personal accounts in the Gemini app only get Flash-Lite, while AI Plus loses Pro (see the Gemini Deep Dive above). Anthropic: Fable 5.1 (9/1), Opus 5.5 (9/22), Sonnet 5.5 (9/28). VWD's position is unchanged: Claude stays the workhorse (logical rigor, enterprise compliance, and a Cowork/Code ecosystem that matches how we work), with Gemini for native video/audio and work inside Google Workspace. Don't switch tools over a headline — the cost of changing platforms dwarfs the gap between models. Sharpening your prompting and verification habits pays far more.
Anthropic 工程师 Thariq Shihipar 发布的 Fable 5 实战方法论,3 天 200 万+ 阅读。核心观点:用好强模型的瓶颈不在模型,而在你自己的"未知"——把未知分为四象限(已知的已知/已知的未知/未知的已知/未知的未知),并给出一套发现它们的技巧,例如开工前先让 Claude 做一轮"盲区扫描"(blind spot pass),把你不知道该问的问题先列出来。与站内 如何写提示词 高度互补,推荐进阶员工阅读。
A hands-on Fable 5 methodology by Anthropic engineer Thariq Shihipar — 2M+ views in 3 days. Core idea: the bottleneck with strong models isn't the model, it's your own unknowns — mapped into four quadrants (known knowns / known unknowns / unknown knowns / unknown unknowns), with concrete techniques to surface them, e.g. asking Claude for a "blind spot pass" before starting unfamiliar work so it lists the questions you didn't know to ask. Pairs well with our Prompting Guide; recommended for advanced users.
x.com/trq212 — A Field Guide to Fable: Finding Your Unknowns
迄今最强的中端 agentic 模型,能力逼近 Opus 档位而价格低得多。原定 2026/9/1 涨到 $3/$15 的计划已于 2026/8/10 被取消,现行标准价仍为输入 $2、输出 $10 /百万 token。详见上方折叠的原始专卡。对 VWD:9/28 起日常默认换成 Sonnet 5.5(同价、更快、更省)。
The strongest mid-tier agentic model yet — near the Opus tier at a much lower price. A rise to $3/$15 was scheduled for 2026/9/1, but Anthropic cancelled it on 2026/8/10; standard pricing stays at $2 in / $10 out per M tokens. See the collapsed original card above. For VWD: from 9/28 the daily default is Sonnet 5.5 (same price, faster, cheaper per task).
面向科研与制药的"研究工作台"——汇集 60+ 科学数据库、把分散数据放到一处处理,结果可审计、可复跑。它不是新模型,而是跑在现有 Claude 模型上的应用。VWD 日常用不到,了解趋势即可:Anthropic 正把能力向更多专业领域延伸。
A "research workbench" for science and pharma — pulls together 60+ scientific databases into one place, with auditable, re-runnable results. It's not a new model, but an app running on existing Claude models. VWD won't use it day-to-day — just awareness that Anthropic is expanding into specialist domains.
Sonnet 4.5 已宣布弃用,Claude API 上 2026 年 11 月 30 日正式下线。日常用 claude.ai 的同事不受影响。IT / 开发同事注意:内部脚本和集成里如果还写着 claude-sonnet-4-5,11/30 前换成 claude-sonnet-5-5(先看上方的 Sonnet 5.5 迁移指南)。
Sonnet 4.5 is deprecated and will be retired from the Claude API on November 30, 2026. If you only use claude.ai, nothing changes for you. For IT / developers: any internal script or integration still calling claude-sonnet-4-5 needs to move to claude-sonnet-5-5 before 11/30 (read the Sonnet 5.5 migration guide above first).
Opus 4.8(2026/5 起的默认旗舰,主打"诚实与可靠")已于 2026/7/24 被 Opus 5 取代。Opus 5 又于 2026/9/22 被 Opus 5.5 接替。如果你在旧培训材料、旧文档里看到"用 Opus 4.8"或"用 Opus 5",一律理解为现在用 Opus 5.5——更强、更快,价格还更低。
Opus 4.8 — the default flagship from May 2026, built around honesty and reliability — was replaced by Opus 5 on July 24, 2026, and Opus 5 was in turn succeeded by Opus 5.5 on September 22. Wherever older training material or documents say "use Opus 4.8" or "use Opus 5," read it as use Opus 5.5: stronger, faster, and cheaper.
💡 实用技巧💡 Practical tips
- 选型口诀(2026/9 月底版):日常 → Sonnet 5.5;难任务(受力计算、合规审查、写代码)以及跑几小时的大任务 → Opus 5.5;高频简单批量 → Haiku 4.5。Fable 5.1 基本不用开——Opus 5.5 已在官方基准上超过它、价格只要四成。看到"Opus 4.8 / Opus 5"就换成 Opus 5.5,看到"Sonnet 5"就换成 Sonnet 5.5。Picking a model (late Sept 2026): daily → Sonnet 5.5; hard work (structural calcs, compliance review, coding) and multi-hour jobs → Opus 5.5; high-volume simple → Haiku 4.5. You'll rarely need Fable 5.1 — Opus 5.5 beats it on Anthropic's benchmarks at 40% of the price. Anywhere you see "Opus 4.8 / Opus 5," read Opus 5.5; anywhere you see "Sonnet 5," read Sonnet 5.5.
- 让它"想得更深":Sonnet 5.5 默认自适应思考,Claude 应用里努力档默认是 medium;难任务可要求更深推理("一步步想清楚、自己核一遍再答"),求快求省时让它直接给结论。Make it think harder: Sonnet 5.5 thinks adaptively by default, and effort defaults to medium in the Claude apps; for hard tasks ask it to reason step by step and check its own work first, and for speed/cost ask for the answer directly.
- 用截图、图表与图纸:Sonnet 5.5 读图表的能力比 Sonnet 5 强得多(图表识别基准 61.6% 对 15.6%),直接贴报表截图、柱状图、CAD 图、报价单让它读——比手敲数据更快更准。读出来的数字仍要抽查,见 验证 AI 输出。Use screenshots, charts & drawings: Sonnet 5.5 reads charts far better than Sonnet 5 (61.6% vs 15.6% on a chart-reading benchmark) — paste report screenshots, bar charts, CAD images, and quotes directly instead of retyping. Still spot-check the numbers it reads; see Fact-Check AI.
- 成本心里有数:Sonnet 5 / 5.5 用的新分词器,同样文本比 Sonnet 4.6 约多 30% token;好在 5.5 完成同样的活用的 token 少得多。长文档/长对话仍要留意上下文占用,必要时用
/compact或拆分。Mind the cost: Sonnet 5 / 5.5 use a newer tokenizer that counts ~30% more tokens than Sonnet 4.6 for the same text — though 5.5 needs far fewer tokens to finish the same job. On long docs/chats, still watch context use; use/compactor split when needed. - 记忆是共享的:从 2026/8/25 起,聊天里让 Claude "记住"的内容在 Cowork 里也生效。好处是不用重复交代背景,代价是别让它记住红线内容——在 Settings → Memory 里可以逐条查看和删除。Memory is shared now: since Aug 25, 2026, anything you ask Claude to remember in chat also applies in Cowork. That saves re-explaining context — but it means don't have it remember redline material. Review and delete entries under Settings → Memory.
- 先学好基本功:写好提示词、验证 AI 输出永远比换模型更重要——见 如何写提示词 与 验证 AI 输出。Fundamentals first: good prompting and fact-checking matter more than switching models — see Prompting Guide and Fact-Check AI.
VWD AI Guide v3.9 | VWD | 2026 | Created by IT depart. | 有问题发邮件到Questions? Email it@valuewds.com