```markdown
## 总体摘要
此次文档更新主要包含两个重点:一是 **Claude Code on the Web 现已支持 GitHub Enterprise Server (GHES)**,二是**移除了 Claude Sonnet 4.5 和 Sonnet 4 的 100 万 token 上下文窗口支持**(回退至 200k)。
## 关键变更主题
* **GitHub Enterprise Server (GHES) 支持**
* `claude-code-on-the-web.md`: 明确 Teams 和 Enterprise 计划现在支持自托管的 GitHub Enterprise Server。
* `network-config.md`: 新增了关于防火墙配置的指南,要求允许 Anthropic API IP 地址以便基础设施访问 GHES。
* `plugin-marketplaces.md`: 推荐使用正则匹配方式来配置 GHES 的内部插件市场。
* **模型上下文窗口策略调整 (重要变更)**
* **移除 Beta 功能**: 移除了 Claude Sonnet 4.5 和 Sonnet 4 的 1M token 上下文窗口 Beta 支持。
* **当前限制**: 现在**只有** Claude Opus 4.6 和 Sonnet 4.6 支持 1M token 上下文窗口,其他模型(包括 Sonnet 4.5/4)上限为 200k token。
* **文档清理**: 移除了所有关于 `context-1m-2025-08-07` Beta header 的提及、长上下文的溢价计费说明以及相关的 Service Tier 计费倍率说明。
## 影响级别
**高**
**理由**: 虽然新增了 GHES 支持是积极的变化,但移除 Sonnet 4.5 和 Sonnet 4 的 1M 上下文窗口能力属于**功能性回退**。正在使用这些模型的 Beta header (`context-1m-2025-08-07`) 处理超过 200k token 数据的开发者将面临行为变更(请求将被截断或无法发送)。
## 开发者行动项
1. **移除 Beta Header**: 如果你正在使用 `context-1m-2025-08-07` header 配合 Sonnet 4.5 或 Sonnet 4,请立即从代码中移除。
2. **迁移模型**: 如果你的应用需要超过 200k token 的上下文,必须迁移至 **Claude Opus 4.6** 或 **Sonnet 4.6**。
3. **更新计费预期**: 针对 Sonnet 4.5/4 的超过 200k token 的高额计费逻辑已不复存在,但需注意 Opus 4.6/Sonnet 4.6 的 1M 窗口按标准费率计费。
4. **配置防火墙 (如适用)**: 如果你是 GHES 用户且使用 Claude Code on the Web,请将 Anthropic API IP 地址加入白名单。
5. **审查插件配置**: 使用 GHES 内部 Git 服务器时,更新插件市场的配置以使用推荐的正则模式。
```
💭 Model Reasoning
1. **Analyze the Request:**
* **Role:** World-class programmer analyzing documentation changes.
* **Output Format:** Markdown, concise, insightful, Chinese language.
* **Input:** 11 pages of diffs.
* **Task:** Batch analysis summarizing changes, themes, impact, and action items.
2. **Analyze the Diffs (Page by Page):**
* **Page 1: `claude-code-on-the-web.md`**
* *Change:* Added support for self-hosted GitHub Enterprise Server (GHES) in "Platform restrictions" section. (Teams and Enterprise plans).
* *Significance:* Expands platform compatibility for Claude Code on the Web.
* **Page 2: `network-config.md`**
* *Change:* Added instruction to allowlist Anthropic API IP addresses for self-hosted GitHub Enterprise Server instances behind firewalls.
* *Significance:* Network configuration guidance for GHES users.
* **Page 3: `plugin-marketplaces.md`**
* *Change:* Added a recommendation to use regex pattern matching for internal git servers, specifically mentioning GitHub Enterprise Server (GHES).
* *Significance:* Best practice guidance for plugin marketplace configuration.
* **Page 4: `about-claude/models/overview.md`**
* *Change:* Updated the "Context window" column in the models table. Previously, Sonnet 4.5 and 4 had 1M (or 200k) options. Now, only Opus 4.6 and Sonnet 4.6 have 200k listed (Wait, looking closer at the diff).
* *Deep Dive:* The diff shows the table row changing from mentioning 1M for Opus/Sonnet to just "200k tokens" for all listed rows (Opus 4.6, Sonnet 4.6, Sonnet 4.5, Sonnet 4, Haiku 4.5). Wait, looking at the table structure:
* Opus 4.6: "200k tokens" (previously 1M or 200k).
* Sonnet 4.6: "200k tokens".
* Sonnet 4.5: "200k tokens".
* Sonnet 4: "200k tokens".
* Haiku 4.5: "200k tokens".
* *Correction:* Looking at the context of other pages, Opus 4.6 and Sonnet 4.6 retain 1M. The table row in this specific diff looks simplified. Let's look at the footer notes. Note 1 was removed (explaining the 1M beta header and long context pricing). Note 2 became Note 1 (explanation of cutoffs). Note 3 became Note 2.
* *Interpretation:* The table itself seems to have been simplified to show the default/base context window (200k), or perhaps the 1M option for Opus 4.6/Sonnet 4.6 isn't explicitly shown in the *tooltip* of the table cell anymore, relying on the text description elsewhere. Or, looking at `context-windows.md` (Page 8), Opus 4.6 and Sonnet 4.6 *do* have 1M.
* *Crucial Observation:* The removal of Note 1 (the one explaining 1M context for Sonnet 4.5/4) indicates Sonnet 4.5 and 4 **lost** their 1M context capability (beta removed/cancelled).
* **Page 5: `about-claude/pricing.md`**
* *Change:* Removed the entire section "Long context pricing" specific to Sonnet 4.5 and Sonnet 4.
* *Change:* Updated "stacking" sentence to remove "long context pricing".
* *Significance:* Confirms that the premium long context pricing for Sonnet 4.5/4 is gone because the feature (1M context) is gone for them. Only Opus 4.6 and Sonnet 4.6 keep 1M at standard pricing (as seen in the remaining text).
* **Page 6: `api/service-tiers.md`**
* *Change:* Removed lines regarding burndown rates for long-context requests (>200k) on Sonnet 4.5 and Sonnet 4.
* *Significance:* Consistent removal of the beta feature logic.
* **Page 7: `build-with-claude/prompt-caching.md`**
* *Change:* Removed "long context pricing" from the list of modifiers that stack with prompt caching.
* *Significance:* Documentation clean-up following the pricing change.
* **Page 8: `build-with-claude/context-windows.md`**
* *Change:* Rewrote the paragraph about context windows.
* *Old:* "Claude Opus 4.6 and Sonnet 4.6 have a 1M-token context window. [Note about Sonnet 4.5/4 beta header]."
* *New:* "Claude Opus 4.6 and Sonnet 4.6 have a 1M-token context window. Other Claude models, including Claude Sonnet 4.5 and Sonnet 4, have a 200k-token context window."
* *Significance:* **MAJOR**. Explicitly states Sonnet 4.5 and 4 are capped at 200k now. The beta 1M context is gone.
* **Page 9, 10, 11 (Cloud Providers: Bedrock, Vertex, Foundry)**
* *Change:* All three pages show similar updates.
* *Pattern:* Changed text from "Opus 4.6, Sonnet 4.6, Sonnet 4.5, and Sonnet 4 have a 1M-token context window..." to "Claude Opus 4.6 and Sonnet 4.6 have a 1M-token context window... Other Claude models... have a 200k-token context window."
* *Significance:* Consistency across cloud platforms. The beta removal is universal.
3. **Synthesize Themes:**
* **Theme 1: GitHub Enterprise Server (GHES) Support.**
* `claude-code-on-the-web.md`: Now supported for Teams/Enterprise.
* `network-config.md`: Firewall allowlisting instructions added.
* `plugin-marketplaces.md`: Recommended regex approach for GHES.
* **Theme 2: Model Context Window Changes (Breaking/Significant).**
* **Retraction:** The beta 1M-token context window for Claude Sonnet 4.5 and Sonnet 4 has been removed/cancelled.
* **Current State:** Only Claude Opus 4.6 and Sonnet 4.6 support 1M tokens.
* **Pricing:** Removed premium long-context pricing for Sonnet 4.5/4 since they no longer support >200k context.
4. **Determine Impact:**
* **GHES Support:** **Low to Medium** impact. Only affects organizations using self-hosted GitHub and Claude Code on the Web. Positive news for them.
* **Context Window Changes:** **High Impact** (Breaking for beta users). Developers relying on the 1M beta header for Sonnet 4.5 or 4 will find their requests truncated or fail if they attempt >200k tokens. This removes a capability that was in beta.
5. **Formulate Action Items:**
* **For Sonnet 4.5/4 users:** Stop using the `context-1m-2025-08-07` beta header. Expect requests >200k tokens to be capped at 200k.
* **Migration:** If you need >200k context, switch to Opus 4.6 or Sonnet 4.6.
* **GHES Admins:** Update firewall rules to allow Anthropic API IPs.
* **GHES Plugin Devs:** Use regex patterns for internal server configuration.
6. **Drafting the Response (in Chinese):**
* *Summary:* Added GHES support for Claude Code on the Web; removed 1M context beta for Sonnet 4.5 and 4, capping them at 200k.
* *Themes:* 1. GHES Integration (Web + Network + Plugins). 2. Context Window Rollback (Sonnet 4.5/4 limited to 200k).
* *Impact:* High (due to feature removal for specific models) / Medium (GHES support is additive). Let's call it "High" because removing a beta feature is technically a breaking change for those using it.
* *Actions:* 1. Migrate to Opus 4.6/Sonnet 4.6 for 1M context. 2. Remove beta header. 3. Configure firewalls for GHES.
7. **Refining the Output (Self-Correction):**
* Make sure to distinguish between *Claude Code* (the IDE agent) and *Claude API*.
* The GHES changes are specific to `claude-code-on-the-web`, not necessarily the general API connecting to GHES (though the network config implies cloud infrastructure connecting to GHES).
* The context window changes apply to API, Bedrock, Vertex, and Foundry.
8. **Final Polish (Markdown Structure):**
* Use bolding for emphasis.
* Keep it concise.
**(Self-Correction during drafting):**
Wait, looking at `models/overview.md`, the context window cell for Opus 4.6 and Sonnet 4.6 *also* shows "200k tokens" in the diff. This seems to be a simplification of the table to show the *default* or perhaps a data error in the diff representation?
However, `context-windows.md` explicitly says: "Claude Opus 4.6 and Sonnet 4.6 have a 1M-token context window." This text was not deleted, only the Sonnet 4.5/4 beta note was deleted.
Therefore: Opus 4.6/Sonnet 4.6 still have 1M. Sonnet 4.5/4 lost it. The table in `overview.md` just got simpler (perhaps hiding the "or 200k" nuance or just showing the row value generically). I will trust the explicit text in `context-windows.md` over the simplified table cell.
Let's stick to the facts derived:
1. GHES support added for Claude Code on the Web.
2. Sonnet 4.5 and 4 are capped at 200k (1M beta removed).
3. Opus 4.6 and Sonnet 4.6 keep 1M.
Action items need to be sharp.
Final check on language: Chinese.
Let's write the response.