發布日期：2026-09-28
當你在終端機呼叫 Claude 處理超大型專案時，現在能用上最新的主力模型。Anthropic API 正式將預設 Sonnet 升級為 Claude Sonnet 5.5（claude-sonnet-5-5）。新模型具備一百萬 token（AI 處理與計費的文字單位）的脈絡容量。輸入每百萬單位兩美元、輸出十美元，快取（把算過的結果暫存起來直接用）讀取為零點二美元。在企業閘道後方使用 Claude 電腦版時，閘道會自動為桌面端標註百萬容量選項。
如果你常常需要 AI 幫忙翻閱專案外的設定檔，現在操作會更靈活。在自動模式（讓 Claude 執行任務時盡量不中途打斷你的運作方式）讀取專案資料夾外的檔案時，提示選項新增了「允許這次，但下次再問」，讓你既能放行眼前的單次讀取，又不會失去後續的把關機會。終端機與 VS Code 上的工作階段在未特別設定權限時，所有方案與供應商都預設以自動模式啟動。你隨時能透過 permissions.defaultMode 設定鍵覆寫預設行為。
想精細微調模型解答時的思考深度，現在不用再被固定按鍵綁死。/effort 的滑桿按鍵現在開放自訂，你可以在 keybindings.json 裡自由綁定 effortSlider:decreaseEffort、increaseEffort 與 toggleUltracode 等動作。Ultracode（可獨立開關的模式）現在能透過 Tab 鍵或 /effort ultracode on 開啟，不再強制綁定 xhigh 思考等級，在任何力度設定下都能持續運作。
查用量或碰到額度上限時，畫面會把花費與等待進度說清楚。企業閘道執行此版或更新版本時，/usage 與 statusLine（終端機底下的狀態列）會顯示花費金額；rate_limits.spend_limit 也新增 used_usd、limit_usd、period 欄位。claude.ai 訂閱者可在 /help 與指令選單找到 /rate-limit-options。/usage、/extra-usage 與 IDE 用量畫面遇到限流或登入遭拒時，會退避再查。限額狀態、倒數與 /usage-credits 選項改放在輸入框下方同一區塊，不再重複洗版；等待時打字，也不會讓再次觸及限額的回合脫離自動接續設定。最高 Max 方案的提示不再叫你升級；/usage-credits 可用時，提示與 /upgrade 都改指向它。
外部工具連不上或要重新登入時，在終端機打 /mcp reconnect all，一次重試全部沒連上的 MCP（讓 AI 連到外部工具的協定）伺服器。接續舊工作階段時，工具若還在連線，呼叫最多等十秒，不會馬上報「找不到工具」。受管設定限定 MCP 伺服器只能由外掛提供時，claude mcp add 會拒絕並說明怎麼辦，不再假報成功。Elicitation、ElicitationResult 掛鉤（事件觸發的小程式）回傳 {"decision":"block"} 時，會拒絕 MCP 的徵詢，不再無視；Chrome 工具少打前綴時，錯誤會指名該叫哪個工具。非互動模式下，被 --allowedTools 或 mcp_tool 掛鉤點名、還在連線的伺服器，第一輪仍等最多兩秒，就算 CLAUDE_CODE_MCP_STARTUP_WAIT_MS 為 0。
身分提供者發憑證而非 client secret 時，支援 private_key_jwt 驗證。telemetry.forward_to 目的地可設 auth: { google: {} }，用閘道的 Google Cloud 憑證，把遙測直傳到它的 OTLP 端點。身分提供者含多群組時，閘道可收 256 KiB 標頭，不再回 431。availableModels 為空或漏列啟動模型且未指定 model 或 enforceAvailableModels 時會警告。ANTHROPIC_FOUNDRY_RESOURCE 不是純資源名稱就拒絕。要求登入（forceLoginMethod 或 forceLoginOrgUUID）卻另設 API key、token 或 apiKeyHelper 時，會列出憑證位置與移除法。
安裝或調整外掛時，/plugin 設定畫面的是／否選項改成二選一，不必手打字；數字欄位遇到無效輸入會拒絕，左右方向鍵也改成切換欄位值，不再跳到別的分頁。新增外掛市集（外掛的集散地）時，若取代了同名但來源不同的舊市集，畫面會說明，也會告訴你怎麼還原。首次安裝時，相依套件的版本範圍對不上而失敗，不會再留下已啟用的紀錄。用 2.39 以前的舊版 git 更新設有 sparsePaths（只抓部分資料夾）的市集時，不會再把能用的本機副本換成空的。Windows 上不存在的外掛 bin/ 資料夾不再被加進 PATH（系統找程式的路徑清單），已有的項目也不會重複加入。公司用受管設定開了 allowManagedPermissionRulesOnly 時，外掛不能再自己預先放行工具，只有 Anthropic 官方來源和受管設定背書的來源例外。
在網路連線不穩或對話過長時，錯誤處理現在會給出更明確的結果。串流（答案一邊生成一邊顯示）損毀時，Claude 會重試或回報回應中斷，不再顯示底層解析錯誤，也不會把 undefined 寫進答案。思考區塊結束後若伺服器過載或出錯，該請求也會重試。對話壓縮後仍超出長度上限時，Claude Code 會再壓縮一次，代價是保留較少近期對話。連線在回應途中斷掉時，重試會與該請求的其他重試共用額度，持續失敗的請求會更早放棄。工作階段指定的模型無法使用且沒有備援模型時，畫面會顯示模型無法使用的說明與了解更多連結。
在狹窄視窗或全螢幕模式下檢視終端機記錄時，介面文字排版不再錯位。在 macOS、Linux 的全螢幕 transcript mode（對話記錄模式）按 [ 將對話寫入 scrollback 時，不會再抹除工作階段上方的終端機輸出；往上捲動歷史記錄時，回覆生成完畢也不會再突兀跳動。/config 與 /plugin 的標籤若超出視窗寬度會整詞換行；/tasks、/copy 與 /hooks 的詳細資訊放得下時會排成一欄，放不下時靠右顯示。Shell 模式按向下鍵不會再誤選隱藏元件，導致無法刪字。/model 清單捲到底時，剩餘模型數不再誤顯示 +1 model。/keybindings 不再把 Backspace、Delete 綁到無效的頁尾動作；重新綁定 footer:close 後，在代理面板按關閉鍵也不會誤輸入 x。
習慣在終端機開啟 Vim 編輯模式的使用者，快速打字後重複變更會更準確。在 SSH 或 tmux 連線下快速輸入，或未使用 bracketed paste（標示貼上內容邊界的方式）時貼上文字，按 . 都能重複上一次的變更。像 cw 後直接按 Esc，沒有輸入文字就按 .，也會正確回到一般模式。在最後一行執行 dd，或在提示框末端執行 yy 後，游標不再落在圖片預留位的起始括號，避免接著按 r 或 x 破壞或刪除圖片。
使用手機或另一台裝置遠端操控終端機時，連線操作的安全性得到補強。當終端機視窗重新取得焦點時，遠端啟用提示會重新啟動防誤觸安全延遲，避免敲擊鍵盤瞬間誤按確認。在尚未信任的資料夾執行 claude remote-control 時，終端機會直接跳出信任確認而非直接結束程式。從行動裝置發送照片並拉回提示框編輯時，圖片不會再遺失；沒有說明文字的照片，也不會再讓游標移動一格。在家目錄啟動後，切換渲染器或更新時，不會再重複彈出信任對話框。
在專案中生成或管理各類工件（AI 產出的網頁、圖表或獨立文件檔案）時，介面與提示現在更清楚。/artifacts 檢視畫面採用了簡潔的分類標籤（全部、我的、已分享）。產出工件頁面時，Claude 會將設計規劃直接寫入頁面本身，並套用你預先命名的標題。資料庫寫入檢視者私有的 data/users/ 子樹時，不再誤稱所有檢視者都看得到；as_level 也新增「view」等級。Artifact 工具收到 claude.ai 聊天或專案連結、聊天中的工件，或單獨的工件 ID 時，會請求正確連結或內容，而非直接中斷。網路分享上的檔案，只有位於以 --add-dir 加入的映射網路磁碟機時才能發布。
當你透過連結引入外部專案設定或累積長期對話記憶時，系統防護現在更周全。透過 symlink（符號連結，指向另一個位置的捷徑檔）將專案外的規則連入 .claude/rules，或將專案外的 .claude 目錄以符號連結接入時，系統都會詢問外部匯入授權，不再默默略過。在讀取自動記憶檔案時，內容中暗藏的不可見字元與偽造標記會被中和過濾。Sonnet 的安全保護標記訊息時，通知會說明原因，並提供修改後重試的選項。若工作階段透過 ANTHROPIC_DEFAULT_OPUS_MODEL 或 modelOverrides 釘選 Opus 模型，且使用 Anthropic API，安全相關的模型切換會由 API 依安全標記類型決定切換模型。
用子代理或自動循環時，有幾處卡點修好了。工作階段經 Claude API 用 Claude Code 不認得的模型 ID（如代理伺服器後方的自訂模型）時，Explore 子代理人會繼承該模型，不再切到 Opus。未提供 SendMessage 的工作階段，不再被要求用它傳訊給其他工作階段。/loop 自行掌握節奏時，每次進度更新與停止結果都會輸出成可見文字。從 macOS 或 Linux 上 Claude 電腦版建立的 worktree（平行副本）啟動 /ultrareview，能上傳工作樹。Agent SDK 遇到 source 格式錯誤的圖片會改放說明文字，格式錯誤的文件區塊也不再讓後續每輪失敗。/recap 從聊天討論串、例行工作或 webhook 轉送時會拒絕；直接在終端機、Claude 應用程式、遠端操控、-p 或 SDK 主機使用仍可。
開發自訂自動化腳本或排查問題時，日誌現在會留下更多有用線索。hook（掛鉤，特定事件發生時被觸發的小程式）失敗時，即使同時寫了正常輸出，偵錯日誌仍會記錄標準錯誤輸出；沒有輸出也會記下狀態碼。Linux 沙盒中的 Bash 若從禁止寫入、且含禁止讀取目錄的工作目錄啟動，不再啟動失敗。監控事件列會顯示事件印出的內容，也不再重複未變的等待訊息。工作流程工具加強了非同步腳本掛鉤錯誤的隔離防護；啟動時只建置設定檔用到的設定架構，縮減記憶體用量並加快啟動。/claude-api 的 hillclimb 不再把評估輪次花在微小到無法測出的提示詞改寫；你在 report.html 旁額外要求的頁面，會產成不連網的單一本機檔案。
在 VS Code 編輯器中使用 Claude Code 時，可以選擇在每則提問與回覆上方顯示時間；日期分隔線只在換日時出現。「顯示訊息時間戳記」設定預設關閉。外掛管理清單的各列會顯示載入錯誤與備註；彈窗可停用、解除安裝外掛，或複製錯誤。思考滑桿下方新增 Ultracode 獨立開關，模型標籤在任何力度下都能顯示啟用狀態。輸入 /model 按 Enter 會開啟模型選擇清單；你自訂的 /mcp、/config、/settings 指令不再被擴充套件的對話框遮蔽。路徑含非 ASCII 字元、空格或括號的檔案連結也能正常點開。
當你在 VS Code 中重啟擴充套件或進行多代理人協作時，工作狀態不會再意外中斷。從 Memory 對話框執行 Reload Claude 重新載入 Claude 時，會先確保編輯中的檔案儲存完畢；重啟擴充套件後，對話頁籤會自動重新連回原先的對話。若還原的頁籤已被其他行程開啟，會主動跳出詢問。焦點檢視中展開的區塊，不會在子代理人工作時或第一個步驟被裁出畫面時自行閉合。來自其他代理人、工作階段或頻道的訊息在重新整理後會完整保留，未標記發送者的代理訊息也不會顯示為原始 XML 標籤。沒有回合在執行時按 Escape 鍵，背景代理人亦會穩定繼續運作。
在特定雲端平台使用 VS Code，或對話夾帶大型文字檔時，回報與壓縮處理更可靠。在 Vertex、Bedrock 或 Foundry 等雲端環境下送出 /feedback 時，回報內容會儲存在本機電腦中。當對話夾帶大型純文字檔案時，對話壓縮機制能妥善處理龐大內容，不再跳出提示過長的錯誤。重新載入視窗後，登入程序不會再因等待 Python 擴充套件而延遲。外掛安裝連結不會再覆寫同名的現有市集。claudeCode.environmentVariables 設定中的 CLAUDE_CONFIG_DIR 只有設為絕對路徑才會生效，並會傳遞給接續對話的終端機。
在 Slack 頻道中標記 @Claude 進行團隊協作時，機器人的回應與狀態通知更清晰。單一討論串、頻道預設或私訊都可選「Opus (latest)」等模型系列，跟隨該系列最新版本。環境拒絕啟動時，Claude 會發布一次聯絡管理員的通知，被再次提及時會重試；未連結帳號的成員每次標記都會收到私訊登入提示。自架環境會說明執行器正在啟動、稍後重試，或不會啟動。新增頻道管理員時，若無法確認頻道的 Slack 工作區已連到組織，錯誤訊息會說清楚原因；新增儲存庫時，若目前的 GitHub 登入無法確認你有管理員權限，頁面會要求登入 GitHub。GitHub Enterprise 主機名稱含底線時，儲存庫預覽與建立 PR 按鈕也能顯示。
企業管理員在後台掌握團隊支出與審查程式碼時，數據檢視更完整。分析預測圖表中現在會列出計入組織總額度的花費與使用比例。管理設定中的「立即通知成員」按一下就能觸及 Enterprise Grid 中組織認領的每個工作區，也能觸及大型工作區的更多成員。頻道存取清單會列出自動關聯的連接器、儲存庫與外掛，以及各自來源。離線時使用雲端例行工作的「編輯」或「複製」控制項，畫面會提示離線，而非說仍在載入。Code Review 遇到 PR 上有其 GitHub App 名下尚未送出的審查時，會先重試發布已完成的審查，不再直接放棄。
When you call Claude in the terminal to work on massive projects, you can now use the latest flagship model. The Anthropic API officially upgrades the default Sonnet to Claude Sonnet 5.5 (claude-sonnet-5-5). The new model features a context window of one million tokens (text units for AI processing and billing). Pricing is two dollars per million tokens for input, ten dollars for output, and zero point two dollars for cache reads. When using Claude for Desktop behind an enterprise gateway, the gateway automatically annotates the one-million context option for desktop clients.
If you often need AI to inspect configuration files outside your project, operations are now more flexible. When reading files outside the project directory in auto mode (an execution mode where Claude runs tasks while minimizing interruptions), a new prompt option "Allow once, ask again next time" lets you permit a single read without losing future gatekeeping control. Sessions in the terminal and VS Code now start in auto mode by default across all plans and providers when permissions are unconfigured. You can override this default behavior at any time using the permissions.defaultMode configuration key.
When fine-tuning the model's reasoning depth, you are no longer locked into fixed key combinations. Slider keys for /effort are now customizable, allowing you to freely bind actions like effortSlider:decreaseEffort, increaseEffort, and toggleUltracode in keybindings.json. Ultracode (an independently toggleable mode) can now be enabled via the Tab key or /effort ultracode on. It is no longer forced to the xhigh reasoning level and continues operating under any effort setting.
When checking usage or hitting quota limits, the interface now clearly explains spending and waiting progress. When the enterprise gateway runs this version or newer, /usage and statusLine (the bar displayed at the bottom of the terminal) show spend amounts; rate_limits.spend_limit also adds used_usd, limit_usd, and period fields. Subscribers on claude.ai can find /rate-limit-options in /help and the command menu. When encountering rate limits or rejected logins, /usage, /extra-usage, and IDE usage views back off before retrying. Rate limit status, countdown timers, and the /usage-credits option are consolidated into a single block beneath the prompt input, eliminating repeated screen clutter. A turn that hits limits again after you type during a wait remains governed by the auto-continuation setting. Prompts for the top Max tier no longer suggest upgrading, and when /usage-credits is available, both prompts and /upgrade redirect to it.
If an external tool you connected to Claude Code can't connect or asks you to sign in again, run /mcp reconnect all in the terminal to retry every MCP (a protocol that lets AI reach external tools) server that isn't connected, all at once. When you resume an old conversation and a tool is still connecting, the call no longer fails right away with "No such tool available" but waits up to ten seconds. If your organization's managed settings only allow plugins to provide MCP servers, claude mcp add no longer reports success while saving a server that never loads; it refuses and explains what to do. When Elicitation or ElicitationResult hooks (small programs that run when specific events happen) return {"decision":"block"}, the MCP request is now declined instead of ignored, and if a Claude in Chrome tool is called without its prefix, the error names the tool to call. In non-interactive mode (no chat screen), a server that is still connecting and is named by --allowedTools or an mcp_tool hook still gets up to two seconds on the first turn, even when CLAUDE_CODE_MCP_STARTUP_WAIT_MS is 0.
When identity providers issue certificates instead of client secrets, private_key_jwt authentication is now supported. The telemetry.forward_to destination can be configured with auth: { google: {} } to forward telemetry directly to its OTLP endpoint using the gateway's Google Cloud credentials. When identity providers include multiple groups, the gateway can accept 256 KiB headers rather than returning 431. A warning is displayed when availableModels is empty or omits startup models while model or enforceAvailableModels is unspecified. Any ANTHROPIC_FOUNDRY_RESOURCE that is not a pure resource name is rejected. When login is required (forceLoginMethod or forceLoginOrgUUID) but an API key, token, or apiKeyHelper is also configured, credential locations and removal instructions are listed.
When you install or adjust plugins, the yes/no options on the /plugin settings screen are now a two-way choice instead of free text; number fields refuse invalid input, and the left and right arrow keys change a field's value instead of jumping to another tab. When you add a plugin marketplace (a place where plugins are collected) that replaces one with the same name from a different source, the screen now says so and how to undo it. If a first install fails because a dependency's version range can't be met, it no longer leaves the plugin recorded as enabled. On older git (before 2.39), updating a marketplace that uses sparsePaths (fetching only some folders) no longer replaces a working local copy with an empty one. On Windows, plugin bin/ folders that don't exist are no longer added to PATH (the list of places the system looks for programs), and entries already there aren't added twice. When your organization's managed settings turn on allowManagedPermissionRulesOnly, plugins can no longer pre-approve tools on their own, except plugins from an official Anthropic source or one the managed settings vouch for.
During unstable network connections or overly long conversations, error handling now delivers clearer outcomes. When a stream (content displayed simultaneously as it generates) is corrupted, Claude retries or reports an interrupted response instead of showing low-level parsing errors or inserting undefined into the answer. If server overload or errors occur after thinking blocks complete, the request also retries. When a conversation exceeds length limits even after compaction, Claude Code compacts it a second time at the cost of retaining fewer recent turns. If a connection drops mid-response, retries share their budget with other retries for that request, failing fast on persistently broken connections. When a session's specified model is unavailable with no fallback configured, the interface displays an unavailability notice alongside a learn-more link.
When viewing terminal transcripts in narrow windows or fullscreen mode, interface text layouts no longer misalign. Pressing [ to write conversations to scrollback in fullscreen transcript mode (conversation transcript view) on macOS and Linux no longer erases terminal output above the session. Scrolling up through history no longer jumps abruptly when responses finish generating. Labels in /config and /plugin wrap on word boundaries if they exceed window width. Detail columns in /tasks, /copy, and /hooks render as single columns when space permits and align right when space is tight. Pressing the Down arrow in Shell mode no longer mistakenly selects hidden elements to block character deletion. The /model list no longer erroneously displays +1 model when scrolled to the very bottom. /keybindings no longer maps Backspace or Delete to invalid footer actions, and rebinding footer:close no longer accidentally inputs x when pressing the close key on agent panels.
Users accustomed to terminal Vim mode will find repeating changes after rapid typing more accurate. When typing quickly over SSH or tmux sessions, or pasting text without bracketed paste (a mechanism marking paste boundaries), pressing . reliably repeats the previous change. Actions such as pressing Esc right after cw and then pressing . without entering text also return correctly to normal mode. After executing dd on the last line or yy at the end of the prompt input, the cursor no longer lands on the opening bracket of image placeholders, preventing subsequent r or x keystrokes from corrupting or deleting images.
Connection security is strengthened when controlling the terminal remotely from a phone or another device. When the terminal window regains focus, the remote activation prompt restarts the safety delay to prevent accidental confirmation from keystrokes. Running claude remote-control in an untrusted directory directly shows a trust confirmation rather than abruptly exiting. Photos sent from mobile devices and pulled back into the prompt box for editing are no longer lost, and photos without captions no longer move the cursor by one space. After launching in the home directory, switching renderers or updating no longer repeatedly pops up the trust dialog.
When generating or managing artifacts (standalone web pages, charts, or document files generated by AI) in a project, the interface and prompts are now clearer. The /artifacts view adopts clean category tabs for All, Mine, and Shared. When generating an artifact page, Claude writes design plans directly into the page itself and applies your pre-named title. Writing to a viewer-private data/users/ subtree in the database no longer incorrectly claims all viewers can see it, and as_level adds a view tier. When the Artifact tool receives claude.ai chat or project links, artifacts in chat, or standalone artifact IDs, it requests the correct link or content instead of abruptly aborting. Files on network shares can only be published if they reside on mapped network drives added with --add-dir.
System defenses are now more comprehensive when linking external project configurations or building long-term conversation memory. Symlinking external rules into .claude/rules or linking external .claude directories now prompts for external import authorization instead of silently skipping them. When reading auto-memory files, invisible characters and forged tags hidden in content are neutralized and filtered. When Sonnet safety protections flag a message, the notification explains the reason and provides an option to edit and retry. If a session pins Opus models via ANTHROPIC_DEFAULT_OPUS_MODEL or modelOverrides while using the Anthropic API, safety-related model switches are determined by the API based on the safety flag type.
Several bottlenecks have been fixed when using subagents or automated loops. When a session uses a model ID unrecognized by Claude Code via the Claude API (such as custom models behind proxy servers), Explore subagents inherit that model instead of switching to Opus. Sessions that do not provide SendMessage are no longer prompted to use it to message other sessions. When /loop paces itself, every progress update and stop outcome is printed as visible text. Launching /ultrareview from a worktree created by Claude Desktop on macOS or Linux can upload the worktree. When the Agent SDK encounters images with malformed source fields, it replaces them with explanatory text, and malformed document blocks no longer cause subsequent turns to fail. /recap rejects requests forwarded from chat threads, scheduled tasks, or webhooks, but remains fully usable directly in terminals, Claude apps, remote control, -p, or SDK hosts.
Logs now preserve more actionable clues when developing custom automation scripts or troubleshooting issues. When a hook fails, debug logs record standard error output even if normal output was produced simultaneously, and status codes are captured even without output. Running Bash inside Linux sandboxes from a write-restricted working directory that contains read-restricted folders no longer fails to start. The monitor event row displays printed event output and no longer repeats unchanged waiting messages. Workflow tools improve isolation against async script hook failures; startup builds only config schemas referenced by configuration files, reducing memory usage and speeding up launch. /claude-api hillclimb no longer spends evaluation rounds on imperceptibly small prompt rewordings, and an extra page requested alongside report.html is generated as one offline, standalone local file.
When using Claude Code in the VS Code editor, you can opt to show timestamps above each prompt and response, with date dividers appearing only when days change. The Show Message Timestamps setting is disabled by default. Plugin management list rows display loading errors and notes, with popups providing options to disable or uninstall plugins and copy error details. An independent Ultracode toggle is added below the thinking slider, and model badges indicate active states across all effort levels. Typing /model and pressing Enter opens the model selection list, and custom /mcp, /config, and /settings commands are no longer obscured by extension dialogs. File links containing non-ASCII characters, spaces, or parentheses can now be opened normally.
Workflows remain uninterrupted when restarting the extension or collaborating across multiple agents in VS Code. Running Reload Claude from the Memory dialog ensures active file edits are saved first, and conversation tabs automatically reconnect to original chats after the extension restarts. If a restored tab is already opened by another process, a prompt asks how to proceed. Expanded blocks in focus views no longer collapse unexpectedly while subagents are working or when the first step is trimmed from view. Messages from other agents, sessions, or channels remain fully intact after refreshes, and uncredited agent messages no longer display as raw XML tags. Background agents continue running stably when pressing Escape while no turn is actively executing.
Reporting and compaction are more reliable when using VS Code on specific cloud platforms or attaching large text files. Submitting /feedback in cloud environments such as Vertex, Bedrock, or Foundry stores report contents directly on the local machine. When conversations include large plain-text file attachments, conversation compaction handles heavy content smoothly without throwing prompt-too-long errors. The login flow no longer delays upon reloading windows while waiting for the Python extension. Plugin installation links no longer overwrite existing marketplaces sharing the same name. CLAUDE_CONFIG_DIR in claudeCode.environmentVariables takes effect only when set to an absolute path, passing through to terminals in continuing sessions.
Bot responses and status notifications are clearer when tagging @Claude for team collaboration in Slack channels. Individual threads, channel defaults, and direct messages can select model families like Opus (latest) to stay aligned with the newest release in that family. When environments decline to launch, Claude posts a single notice to contact administrators and retries upon subsequent mentions, while unlinked members receive a direct message login prompt on every mention. Self-hosted setups clarify whether runners are starting, should be retried later, or will not start. Adding channel admins provides clear explanations when a channel's Slack workspace cannot be verified as connected to the organization; adding repositories requests a GitHub login if the current session cannot confirm admin permissions. Repository previews and PR creation buttons display properly even when GitHub Enterprise hostnames include underscores.
Enterprise administrators gain a more complete data view when tracking team spending and reviewing code. Forecast analytics charts now break down expenditures and percentage utilization counted against the organization quota. The Notify Members Now control in administrative settings reaches every claimed workspace across an Enterprise Grid in a single click, reaching broader member sets in large workspaces. Channel access listings display automatically linked connectors, repositories, and plugins alongside their respective origins. Interacting with edit or duplicate controls for cloud scheduled tasks while offline prompts an offline notice rather than indicating ongoing loading. When Code Review encounters pending unsubmitted reviews under its GitHub App on a pull request, it retries publishing completed reviews rather than giving up immediately.
Added Claude Sonnet 5.5 (`claude-sonnet-5-5`), now the default Sonnet model on the Anthropic API — 1M context, $2/$10 per Mtok with $0.20/Mtok cache reads
Fixed Claude Desktop behind a Claude apps gateway offering no 1M context option: the gateway now marks each 1M-capable model for Desktop automaticallyAdded a "Yes, but ask again next time" answer to auto mode's prompt before a read outside the working directories, so you can allow that one read and still be asked about later ones
Changed interactive terminal and VS Code sessions to start in auto mode when no permission mode is configured, on every plan and provider; `permissions.defaultMode` still overrides itAdded `effortSlider:decreaseEffort`, `increaseEffort` and `toggleUltracode` keybinding actions, so the `/effort` slider's arrow and Tab keys can be rebound in `keybindings.json`
Changed Ultracode into its own toggle in `/effort` (Tab, or `/effort ultracode [on|off]`): it no longer forces xhigh effort and stays on at any effort levelAdded dollar amounts to the Claude apps gateway spend limit in `/usage` and the status line (for example "$271.40 / $500.00 spent this month") when the gateway runs this version or later; the status line's `rate_limits.spend_limit` also gains `used_usd`, `limit_usd` and `period`
Added `/rate-limit-options` to `/help` and the command menu for claude.ai subscribers, so the usage-limit notices that mention it point to a command you can find
Fixed repeated calls to the plan-usage endpoint after it rate-limits or rejects your login: `/usage`, `/extra-usage` and IDE usage views now back off instead of re-asking
Fixed typing a message during an automatic usage-limit wait taking the wait out of the "Continue automatically at usage limit" setting's control when that turn hit the limit again
Fixed usage-limit warnings suggesting `/upgrade` to users already on the highest Max plan; the warnings and `/upgrade` itself now point at `/usage-credits` when it is available
Improved the usage-limit wait: the limit's state and the countdown with the usage-credits option now show as one block under the prompt, and limit messages no longer repeat the countdownAdded `/mcp reconnect all` in the interactive terminal to retry every MCP server that failed to connect or needs authentication at once
Fixed MCP tool calls in a resumed session failing with "No such tool available" while their server was still connecting; the call now waits up to 10 seconds for the server
Fixed `claude mcp add` reporting success when managed settings restrict MCP servers to plugins; it now refuses and says what to do, instead of saving a server that never loads
Fixed `{"decision":"block"}` returned by Elicitation and ElicitationResult hooks being ignored; it now declines the MCP elicitation, as exit code 2 does
Improved the "No such tool available" error for Claude in Chrome tools called without their prefix: it now names the tool to call
Changed the non-interactive first turn to still wait up to 2s for connecting MCP servers named by `--allowedTools` or an `mcp_tool` hook, even when `CLAUDE_CODE_MCP_STARTUP_WAIT_MS` is `0`Added Claude apps gateway startup warnings when a managed policy's `availableModels` is empty, or leaves out the model Claude Code starts on without setting `model` or `enforceAvailableModels`
Added `auth: { google: {} }` for Claude apps gateway `telemetry.forward_to` destinations, so telemetry can be exported straight to Google Cloud's OTLP endpoint using the gateway's Google Cloud credentials
Added certificate client authentication (`private_key_jwt`) between the Claude apps gateway and its identity provider, for identity providers that issue certificate credentials instead of client secrets
Fixed `ANTHROPIC_FOUNDRY_RESOURCE` being interpolated into the Foundry endpoint host unvalidated; a value that is not a plain resource name is now refused
Fixed the Claude apps gateway answering `431 Request Header Fields Too Large` to every request from a sign-in whose identity provider lists many groups; it now accepts request headers up to 256 KiB
Improved the startup refusal when managed settings require a sign-in (`forceLoginMethod` or `forceLoginOrgUUID`) and an API key, token or `apiKeyHelper` is configured: it now names the credential in use, where it is set, and how to remove itFixed the `/plugin` configure screen: boolean options are now a true/false choice instead of free text, number options refuse invalid input, and ←/→ change an options field instead of switching tabs
Fixed Bash tool failing on Windows with many plugins enabled: plugin `bin/` directories that don't exist are no longer added to PATH, and inherited entries aren't added twice
Fixed `sparsePaths` plugin marketplaces cloning empty and replacing a working local copy on older git (before 2.39), which failed every refresh with "marketplace.json file is no longer present"
Fixed plugins from marketplaces, claude.ai and npm pre-approving their own tools via `allowed-tools` under managed `allowManagedPermissionRulesOnly`; only plugins from an official Anthropic source or a source that managed settings vouch for keep that pre-approval
Fixed a failed first `claude plugin install` leaving the plugin enabled and recorded when a dependency's version range could not be met
Improved `claude plugin marketplace add` to say when it replaces a marketplace already added under the same name from a different source, and how to undo itFixed a damaged response stream showing raw errors such as "JSON Parse error" or "undefined is not an object", or writing the word "undefined" into an answer, instead of being retried or reported as an interrupted response
Fixed an overloaded or server error arriving right after a thinking block ending the turn with an error instead of being retried
Fixed "Prompt is too long" errors that persisted after compacting: when the compacted request is still too long, Claude Code now compacts once more, keeping less of the recent conversation
Fixed a session whose model is unavailable, with no fallback model left, showing a bare "is currently unavailable" message (or "Something went wrong" in cloud sessions) instead of the model-unavailable notice and its Learn more link
Changed retries after a dropped connection mid-response to share one budget with the rest of the request's retries, so a failing request gives up soonerFixed ↓ in shell mode selecting a hidden background-tasks pill, which stopped Backspace and Ctrl+U from editing the prompt
Fixed fullscreen rendering erasing the terminal output above the session when `[` in transcript mode writes the conversation to scrollback (macOS and Linux)
Fixed fullscreen scroll position jumping to the previous message or to the bottom when a reply finished streaming while scrolled up
Fixed tab bars in dialogs such as `/config` and `/plugin` breaking the title and tab labels mid-word in a narrow terminal; a tab that doesn't fit now moves to the next line whole
Fixed the `/model` picker showing "+1 model" below the list after scrolling to the last model; the count now covers only the models below the visible rows
Fixed `/keybindings` writing Backspace and Delete bindings for a footer action that does nothing into the generated `keybindings.json`
Fixed a rebound agent panel close key (`footer:close`) typing "x" instead of itself on the row of the agent you're viewing
Improved lists such as `/tasks`, `/copy` and `/hooks`: the details after each name now line up in one column when they fit, and otherwise sit at the right edgeFixed vim mode `.` not repeating text typed very fast (for example over ssh or in tmux) or pasted without bracketed paste, and leaving the prompt in INSERT mode after repeating a change with nothing typed (such as `cw` then Esc)
Fixed vim mode leaving the cursor on an image placeholder's opening bracket after `dd` on the last line or `yy` at the end of the prompt, where `r` or `x` would break or delete the imageFixed a key pressed the instant the terminal regained focus answering the Remote Control enable prompt before its short safety delay restarted
Fixed the workspace trust dialog appearing a second time after switching renderers or updating when Claude Code was started in the home directory
Fixed a photo sent from the Claude app over Remote Control being lost when its queued message was pulled back into the terminal prompt to edit, and the cursor moving one character for a photo with no caption
Improved `claude remote-control`: in a folder you haven't trusted yet, it now asks for workspace trust on the terminal instead of exitingFixed artifact database write results telling Claude that every viewer sees a write to a viewer's private `data/users/` subtree, and added a "view" level to `as_level`
Improved artifact pages: Claude writes its design plan into the page instead of the reply, and uses the name you already gave something as the page title
Improved the Artifact tool so that when Claude is given a claude.ai chat or project link, an artifact from a chat, or an artifact id on its own, it asks for the right link or the content instead of stopping
Changed `/artifacts` to show its filter tabs beside the title with one-word labels (All, Mine, Shared), using the same tab bar as `/config` and `/plugin`
Changed artifact publishing to refuse a file on a network share (a `\\host\share` path or a `/net` automount) unless it is on a mapped network drive added with `--add-dir`Fixed rules symlinked into `.claude/rules` from outside the project being skipped without ever showing the external-imports approval prompt; a `.claude` directory symlinked from outside the project now asks for the same approval
Improved auto-memory loading: invisible characters and tags that imitate Claude Code's own markup are neutralized in `MEMORY.md` and recalled memory notes before they reach Claude
Changed the notice shown when a Sonnet model's safeguards flag a message to explain why it happened and to offer editing and retrying
Changed safety-related model switches in sessions that pin an Opus model with `ANTHROPIC_DEFAULT_OPUS_MODEL` or `modelOverrides`: on the Anthropic API, the API now picks the model to switch to for each kind of flag, not the pinned modelFixed Agent SDK sessions crashing when a user message contains an image with a malformed `source`, and failing on every later turn after a malformed document block; a malformed image is now replaced with an explanatory note
Fixed sessions launched without the `SendMessage` tool (such as by Claude Desktop) still being told to message other sessions with it
Fixed the Explore subagent switching to Opus on the Claude API when the session runs a model ID Claude Code doesn't recognize, such as a custom model behind a proxy; Explore now inherits that model
Fixed `/loop` status updates in self-paced mode often not being shown because Claude wrote them only in its reasoning; Claude now writes each update, and the outcome when the loop stops, as visible text
Fixed `/ultrareview` failing to upload the working tree when started from a git worktree that the Claude desktop app created on macOS or Linux
Changed `/recap` to decline with a short notice when it arrives relayed from a chat thread (your own included) or from a routine or webhook; typed in the terminal, the Claude apps, Remote Control, `-p` or an SDK host, it runs as beforeFixed the debug log dropping a failed hook's stderr when the hook also wrote to stdout, and logging nothing for a failed hook with no output; failed hooks now also log their status code
Fixed sandboxed Bash commands failing to start on Linux when the working directory is write-denied and contains a read-denied directory
Improved Monitor event rows to show what each event printed instead of repeating the description, and stopped repeating an unchanged "Waiting for N … to finish" line after every event
Improved Workflow tool sandbox hardening for errors thrown by async script hooks
Improved startup time and memory use by building only the parts of the settings schema that your settings files actually use
Improved `/claude-api`: `hillclimb` no longer spends rounds on prompt rewordings too small for the eval to measure, and an extra page you ask for beside `report.html` is built as one local file that loads nothing from the network[VSCode] Added an optional time above each prompt and response, with a date line where the day changes (Claude Code: Show Message Timestamps setting, off by default)
[VSCode] Added plugin load errors and notes to the Manage plugins rows, with a popup to disable, uninstall or copy the error
[VSCode] Added an Ultracode on/off switch under the Effort slider, replacing the slider's Ultracode stop; the model pill shows "· Ultracode" at any effort level
[VSCode] Fixed typing `/model` and Enter printing usage text into the chat instead of opening the model selector
[VSCode] Fixed a user's own `/mcp`, `/config` or `/settings` command being shadowed by the extension's dialog
[VSCode] Fixed chat links to files with non-ASCII characters, spaces or brackets in their path not opening[VSCode] Fixed Reload Claude from the Memory dialog restarting before an edited file was saved
[VSCode] Fixed a restored tab opening a conversation another Claude process still has open; it now asks first
[VSCode] Fixed Focus view sections you expanded closing on their own while a sub-agent is working or when the section's first step is trimmed from view
[VSCode] Fixed Claude Code tabs that stopped responding after Restart Extensions: they now reopen on their conversation
[VSCode] Fixed a message from another agent with no recorded sender showing as raw XML in the chat
[VSCode] Fixed messages from other agents, sessions or channels disappearing after a reload
[VSCode] Fixed Escape stopping every background agent when no turn was running[VSCode] Fixed `/feedback` on Vertex, Bedrock and Foundry being refused after you pressed Send; the report is now saved on this computer, as the terminal does
[VSCode] Fixed sign-in waiting up to a minute for the Python extension after a window reload
[VSCode] Fixed plugin install links replacing a marketplace you already have that uses the same name
[VSCode] Fixed "Prompt is too long" errors after compaction when a large text file is attached to a message
[VSCode] Changed `CLAUDE_CONFIG_DIR` in the `claudeCode.environmentVariables` setting to apply only when it is an absolute path, and passed it to terminals that continue the chat[Claude Tag] Added model family choices such as "Opus (latest)" for a thread, a channel default or your DM, so the choice follows the newest model in that family
[Claude Tag] Fixed the earlier Claude in Slack app's progress card and link previews omitting the repository and Create PR button when a GitHub Enterprise host name contains an underscore
[Claude Tag] Fixed Claude staying silent in a channel whose environment declines to start it; it now posts one notice asking you to contact an admin, and retries when @-mentioned
[Claude Tag] Changed Claude to post its private sign-in notice at every @mention from someone who hasn't connected their Claude account, instead of going quiet after the first
[Claude Tag] Improved Claude's wait notice on self-hosted environments with on-demand runners: it now says whether a runner is starting, a start will be retried, or no runner will start
[Claude Tag] Improved the error shown when adding a channel manager fails because the channel's Slack workspace can't be confirmed as connected to your organization
[Claude Tag] Improved adding repositories as a channel manager: when your GitHub sign-in can't confirm you're a repository admin, the page asks you to sign in with GitHub[Cloud sessions] Fixed a routine's Edit and Duplicate controls saying the routine was still loading while you were offline; they now tell you you're offline
[Claude Tag] Added the spend that counts toward your organization-wide limit to the analytics spend projection chart, with how much of the limit is used
[Claude Tag] Improved "Notify members now" in admin settings: one press reaches every workspace your organization claimed in an Enterprise Grid, and more members in large workspaces
[Claude Tag] Improved a channel's access lists in admin settings to show the connectors, repositories and plugins an auto-join pattern attaches, and where each comes from
[Code Review] Fixed Code Review giving up without posting a finished review when an unsubmitted review under its GitHub App was open on the pull request; it now retries the post first