Claude docs changes for August 13th, 2026 [diff]

Executive Summary

• The Sessions API (Managed Agents) gains a major spend-control and observability upgrade: a new budget field lets you cap session cost (with a budget_reached stop reason and session.usage streaming event), sessions can pin an inference_geo, coordinators can add a platform "advisor" roster entry, and usage objects now report active_seconds, list_cost, and server_tool_use — see the new Session Budgets guide.
• Claude Code skills can now sync automatically from your claude.ai account into ~/.claude/skills/synced/ via CLAUDE_CODE_SYNC_SKILLS, with new conflict-resolution and safety rules for synced skills.
• New Agent SDK controls to bound subagent trees: CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH, CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS, and a per-query maxBudgetUsd spend cap.
• A new anthropic-workspace-id response header identifies which workspace a request resolved to, and the Default Workspace now has a real wrkspc_-prefixed ID instead of none.
• New Claude Code versions 2.1.229 and 2.1.231 fix streaming display glitches, several crash bugs, and MCP OAuth redirect issues.

New Claude Code versions

2.1.229

New features

• Documented claude remote-control --continue for resuming the most recent Remote Control session
• Added server-supplied Claude Code hook support for self-hosted runner sessions, matching managed-environment behavior
• Added plugin marketplace command sources: a local command (e.g. an IDE) prints the plugin directory, re-resolved each session and applied without a restart
• ListAgents now marks disconnected Remote Control sessions as offline and labels cloud sessions as cloud
• [VSCode] Added session groups in the sidebar — right-click to create, rename, or delete; Cmd/Ctrl- or Shift-click to move several sessions at once

Existing feature improvements

• Added SSE keepalive pings to gateway streaming responses during long thinking pauses, preventing idle-timeout disconnects on Vertex and Bedrock upstreams
• Improved workflow fan-outs to stagger same-prefix sibling agents so subsequent agents read the cached prompt prefix instead of re-paying it
• Improved "prompt is too long" errors to explain why automatic compaction could not recover instead of only suggesting /compact
• Improved sandbox: IPv6 literals in network domain lists are now bracketed, and ambiguous spellings are enforced fail-closed and flagged by /doctor
• Changed /commit-push-pr so git/gh commands with dangerous flags (--force, --amend, --no-verify, etc.) are no longer auto-approved
• [VSCode] "Report a problem" and /bug now open the built-in feedback dialog instead of a retired survey link; made the /btw side-question panel resizable

Major bug fixes

• Fixed long responses partly disappearing while streaming and being printed twice in the terminal
• Fixed a crash to the error screen when a tool call had a non-string glob, file_path, or command value
• Fixed a RangeError crash when a progress bar or markdown table rendered in a very narrow terminal window
• Fixed a crash on Windows when a tool call or message referenced a file by an extended-length or UNC path
• Fixed auto mode failing on every tool call for users who disable the attribution header via CLAUDE_CODE_ATTRIBUTION_HEADER
• Fixed MCP OAuth with strict authorization servers by using 127.0.0.1 instead of localhost in the redirect URI
• Fixed Remote Control clients showing a stuck working spinner after a slash command typed in the laptop terminal
• Fixed the Claude Code Review workflow generated by /install-github-app completing without posting its review on the pull request
• Fixed multi-second UI stalls after editing a file with thousands of IDE diagnostics while the IDE extension is connected
• Fixed dynamic workflows inside CPU-limited containers using the host machine's core count instead of the container's CPU limit
• Fixed conversations whose messages alone exceed the API's 32 MB request limit retrying compaction endlessly; they now fail once with a clear message
• Fixed self-hosted runner and other remote sessions exiting at startup when managed-mcp.json is deployed and the server delivers MCP servers
• Fixed self-hosted runner repository preparation hanging on a Git Credential Manager prompt

2.1.231

Major bug fixes

• Fixed MCP OAuth sign-in failing with a redirect URI mismatch for servers that use a pre-registered OAuth client, such as Slack

Claude Code changes

Changed documents

Admin Setup [Source]

• The "Login enforcement" row is corrected to say forceLoginOrgUUID is only verified for claude.ai account logins on terminal/VS Code/Agent SDK, and is not checked for Claude Console logins or gateway sign-in — narrower than the previous wording. [line 79] [Source]

Advisor [Source]

• Fable 5 is now offered as an advisor model: /advisor fable and --advisor fable now work, and the pairing table/aliases no longer describe Fable as rejected. [lines 66-86] [Source]
• New requirement: the advisor tool is gated behind feature-flag fetching, so DISABLE_TELEMETRY (or similar) keeps the advisor off even on a supported model. [line 117] [Source]

Agent SDK - Agent Loop [Source]

• Clarifies session_store/sessionStore mirrors transcripts so another host (not "any host") can resume them, aligning with the working-directory-scoped lookup key. [line 268] [Source]

Agent SDK - Cost Tracking [Source]

• New guidance: to bound how much subagents add to total_cost_usd, set the depth, concurrency, and spend limits documented on the subagents page. [line 116] [Source]

Agent SDK - MCP [Source]

• New MCP_TOOL_TIMEOUT env var documented as separate from MCP_TIMEOUT: MCP_TIMEOUT bounds only the connection attempt, while MCP_TOOL_TIMEOUT controls how long a running tool call may take. [line 858] [Source]
• Clarified the init message's tools array only lists tools for servers connected by the time the message is emitted plus cached-tool-list servers. [line 206] [Source]

Agent SDK - Plugins [Source]

• Plugin skills can now be invoked directly with /plugin-name:skill-name (previously just /skill-name), and commands/ is reframed as a supported location for skills as flat Markdown files rather than a legacy format. [lines 5-276] [Source]

Agent SDK - Python [Source]

• New output_tokens_details.thinking_tokens field on the Agent tool's usage, reporting thinking tokens in the subagent's final request (Python SDK v0.2.136+ / Claude Code v2.1.228+). [lines 2342-2391] [Source]
• Clarifies usage/totalTokens on the Agent tool result reflect only the subagent's final API request, not the whole run. [line 2389] [Source]

Agent SDK - Session Storage [Source]

• Clarifies resuming via a session store requires the new host to run from a working directory matching the original run's, since projectKey encodes the working directory. [lines 3-84] [Source]

Agent SDK - Sessions [Source]

• "Resume across hosts" now lists passing a session store as a first-class option, noting the store lookup key derives from the working directory. [line 378] [Source]

Agent SDK - Slash Commands [Source]

• Reframes custom commands as merged into skills rather than deprecated: .claude/commands/ files and .claude/skills/<name>/SKILL.md both create /name and work equivalently. [line 205] [Source]

Agent SDK - Subagents [Source]

• New section "Cap subagent depth, concurrency, and spend" documents CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH (default 3), CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS (default 20), and the new maxBudgetUsd/max_budget_usd query option, including behavior at each limit. [line 589] [Source]
• New "Run Opus 5 with subagents" section: Opus 5 delegates to subagents more readily, and Claude Code auto-adds a delegation-restraint line to its system prompt when using the claude_code preset with Opus 5. [line 670] [Source]

Agent SDK - TypeScript [Source]

• New output_tokens_details.thinking_tokens field on AgentOutput.usage (TypeScript SDK v0.3.228+ / Claude Code v2.1.228+). [lines 2913-2956] [Source]

Agent View [Source]

• New v2.1.227 changelog entry: deleting a session now keeps the session and its worktree if another live session is still running inside that worktree, showing not deleted in agent view and kept <id> from claude rm. [line 701] [Source]

Authentication [Source]

• Significant clarification: forceLoginOrgUUID only restricts/verifies claude.ai account logins. For Claude Console logins it merely pre-selects the org and does not enforce which org the resulting credential belongs to; admins must set forceLoginMethod: "claudeai" to actually force claude.ai sign-in. [lines 19-123] [Source]

Best Practices [Source]

• Corrects auto mode's non-interactive behavior: when the classifier repeatedly blocks actions in a -p run, Claude Code no longer aborts the run. [line 499] [Source]

Claude Apps Gateway - Config [Source]

• The gateway now also rejects requests with a missing/empty model value, in addition to the existing non-string-model check (gateway v2.1.228+). [line 442] [Source]
• Two new Claude Desktop policy keys: chatTabEnabled and chatAdvancedFileAnalysisEnabled (lets Claude run code in a local sandbox to analyze attached files); both require gateway v2.1.227+. [lines 554-561] [Source]

Claude Apps Gateway - Deploy [Source]

• As of v2.1.227, claude gateway's host process sends nothing to Anthropic, removing the need to set CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC for strict-egress deployments on current releases. [line 190] [Source]

Claude Directory [Source]

• A project's auto memory directory is now excluded from the age-based retention sweep and only removed once empty for the full retention period — before v2.1.228 the sweep could delete old files inside it as if they were session data. [lines 202-205] [Source]
• Deleting ~/.claude/projects/ now also removes auto memory for every project. [line 285] [Source]

CLI Reference [Source]

• claude import is now documented as unavailable on Bedrock/Vertex/Foundry/Claude Platform on AWS and when feature-flag fetching is off; --advisor now accepts fable. [lines 27-52] [Source]

Commands [Source]

• /advisor now accepts fable; /import gains the same provider/feature-flag unavailability notes; /loop's self-pacing now depends on feature-flag availability. [lines 35-85] [Source]

Cross-Session Messaging [Source]

• While connected to Remote Control, cross-machine messages now show up under the sender's Remote Control name in the recipient's conversation. [line 80] [Source]
• New CLAUDE_CODE_MESSAGING_TOKEN env var (v2.1.228+): a per-session auth token used to verify "own-child" messages when process-based verification isn't available. [lines 153-158] [Source]

Env Vars [Source]

• New "Features that need feature-flag fetching" section enumerates everything that stops working when DISABLE_GROWTHBOOK/DISABLE_TELEMETRY/DO_NOT_TRACK/CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC disable flag fetching. [lines 439-453] [Source]
• New CLAUDE_CODE_MESSAGING_TOKEN variable row documented. [line 275] [Source]
• CLAUDE_CODE_SYNC_SKILLS behavior refined: Claude Code now waits only for the skill list before the first query, downloads finish in the background, and downloaded skills get extra safety rules. [line 330] [Source]

Errors [Source]

• Reworks what happens on auto-mode classifier refusals in non-interactive runs: a background subagent gets an Agent aborted error result while the run continues, whereas the main conversation receives the denial and continues rather than the whole process exiting. [line 302] [Source]

Fullscreen [Source]

• /tui fullscreen relaunch now also carries over the model you last picked with /model. [line 14] [Source]

Headless Mode [Source]

• Clarified that acceptEdits mode also auto-approves the read-only command set, not just filesystem commands like mkdir/touch/mv/cp. [line 222] [Source]

Hooks Reference [Source]

• Removed documentation of the ConfigChange hook input's source and file_path fields. [line 2360] [Source]

Memory [Source]

• Session transcripts are deleted after the retention period, but auto-memory directory files (MEMORY.md and topic files) are excluded from that sweep and persist until explicitly edited or deleted. [line 340] [Source]

Permissions [Source]

• Removed guidance that using only the tool name without parentheses (Read, Edit, or Write) allows all file access. [line 294] [Source]

Plugin Marketplaces [Source]

• Replaced the explicit 4-step version-resolution order with a summary and a pointer to the full order in plugins-reference.md. [line 810] [Source]

Plugins Reference [Source]

• Claude Code only removes a cached plugin/marketplace folder when it holds no directory or symlink, so a development checkout symlinked into the plugin cache is never marked orphaned or removed. [line 698] [Source]

Remote Control [Source]

• New "What connected devices see" section clarifies /resume conversation-switch sync behavior and when the server session gets archived on reconnect. [lines 128-135] [Source]

Sandbox Environments [Source]

• Docker's microVM sandboxing is now called "Docker Sandboxes," is a free standalone product, and no longer requires Docker Desktop. [line 133] [Source]

Scheduled Tasks [Source]

• Turning off feature-flag fetching now also forces the fixed 10-minute /loop schedule / usage-message behavior, previously documented only for third-party gateway platforms. [lines 64-107] [Source]

Self-Hosted Environments [Source]

• Expanded --retire-at release-sequence description: a session whose background tasks finished but whose follow-up turn hasn't started yet is now held before release. [lines 81-83] [Source]

Self-Hosted Environments - Deploy [Source]

• Shutdown-timing guidance restructured; the --retire-at margin must now also account for the new background-task hold described in Runner lifecycle. [lines 236-239] [Source]

Self-Hosted Environments - Reference [Source]

• New SELF_HOSTED_RUNNER_BG_RESULT_GRACE_MS env var (default 30000ms, v2.1.228+): controls how long the runner treats a session as busy after a background task finishes but before the follow-up turn starts. [line 79] [Source]

Settings [Source]

• New merge rules for managed-settings.d/ drop-in files: a later file's fallbackModel chain and extraKnownMarketplaces entries now replace rather than merge with an earlier file's. [lines 102-846] [Source]
• advisorModel now accepts "fable" as a model alias. [line 193] [Source]
• env setting now also ignores CLAUDE_CODE_MESSAGING_TOKEN, since Claude Code exports it itself (v2.1.228+). [line 252] [Source]

Skills [Source]

• Major new feature: skills synced from your claude.ai account. New section documents enabling CLAUDE_CODE_SYNC_SKILLS to download claude.ai-enabled skills into ~/.claude/skills/synced/, plus name-conflict resolution and safety rules restricting ! shell commands from synced skill bodies. [lines 173-228] [Source]

Tools Reference [Source]

• RemoteTrigger and ScheduleWakeup tools are now also unavailable when feature-flag fetching is turned off. [lines 35-37] [Source]

Troubleshoot Install [Source]

• New detailed explanation of how Claude Code locates bash.exe on Windows when CLAUDE_CODE_GIT_BASH_PATH isn't set. [lines 621-626] [Source]

Ultrareview [Source]

• Removed the "free-run period ends" condition — paid usage-credit billing now applies once the three free runs are used, with no separate time-based expiry. [line 91] [Source]

Zero Data Retention [Source]

• Clarified that forceLoginMethod/forceLoginOrgUUID restrict claude.ai logins specifically, referencing new documentation on how Claude Console logins are treated differently. [line 24] [Source]

API changes

New Documents

Home [Source]

This new landing page for the Claude Platform introduces the "Start building with Claude" hub, offering a quick-start code sample and links to the API key and API reference. It contrasts the two ways to build with Claude — direct Messages API access versus the fully managed Claude Managed Agents infrastructure — and walks through a full developer journey (get started, build, evaluate/ship, operate). It also lists the current Claude model family (Fable 5, Opus 5, Sonnet 5, Haiku 4.5) with brief positioning for each, plus links to courses, the cookbook, quickstarts, release notes, and Claude Code.

Session Budgets [Source]

This brand-new page documents the optional budget field on Managed Agents sessions and deployments, a hard spend ceiling priced against public list rates for model tokens, web searches, and session running time. It explains how list cost is measured and enforced between (not mid-) model requests, what happens when a session reaches budget_reached (it pauses/idles rather than terminating, and only "settle" events like tool results are accepted), how to change or remove a budget to resume work, how budgets behave in multiagent sessions and on scheduled deployments, and the specific error conditions that reject budget-related requests.

Changed documents

Admin - MCP Tunnels - Tunnel Certificates - Create [Source]

• Fixed the example curl request body: the ca_certificate_pem value previously used double-escaped \\n sequences, now uses single-escaped \n so copy-pasting produces a correctly newline-formatted PEM certificate. [line 83] [Source]

Agents And Tools - Tool Use - Advisor Tool [Source]

• Added a concrete side-by-side example showing the same advisor request sent with claude-opus-4-8 (returns plaintext advisor_result) versus claude-opus-5 (returns encrypted advisor_redacted_result). [lines 134-166] [Source]
• Clarified multi-turn behavior: with a Claude Opus 5 advisor, the advisor_redacted_result block must be round-tripped verbatim. [line 205] [Source]

Agents And Tools - Tool Use - Code Execution Tool [Source]

• Clarified guidance on tool version selection: use code_execution_20260120 or later for programmatic tool calling or REPL state persistence, versus code_execution_20250825 for basic Bash/file operations. [line 39] [Source]

Agents And Tools - Tool Use - Tool Search Tool [Source]

• Documented a new limit field Claude can set in its tool search input, an integer from 1 to 10,000 (default 5), letting Claude request more or fewer matching tools than the previous fixed default. [line 217] [Source]

API Overview [Source]

• Response headers section reworked into a table and now documents a new anthropic-workspace-id header (the wrkspc_-prefixed workspace ID the credential resolved to). [line 118] [Source]

Beta [Source]

• New mid-conversation-tool-changes-2026-07-01 beta header value documented, alongside a new session.budget_reached webhook event type. [line 91] [Source]
• New BetaCurrency and BetaMonetaryAmount types added, representing monetary amounts as minor-unit integer strings to avoid float rounding. [line 117] [Source]

Beta - Agents [Source]

• New BetaManagedAgentsAdvisor roster entry type: a platform advisor model the session's primary thread may consult mid-turn, usable in a coordinator's agents array. [line 35] [Source]
• BetaManagedAgentsModelConfig gains a new inference_geo field to select the geographic region for model inference. [line 203] [Source]

Beta - Agents - Archive [Source]

• Response schema now reflects the new inference_geo model config field and the BetaManagedAgentsAdvisor coordinator roster entry type. [line 263] [Source]

Beta - Agents - Create [Source]

• model_config body parameter gains inference_geo, and a new BetaManagedAgentsAdvisorParams roster entry type is available when configuring a coordinator's agents array. [line 329] [Source]

Beta - Agents - List [Source]

• Response schema now reflects the new inference_geo model config field and the BetaManagedAgentsAdvisor coordinator roster entry type. [line 281] [Source]

Beta - Agents - Retrieve [Source]

• Response schema now reflects the new inference_geo model config field and the BetaManagedAgentsAdvisor coordinator roster entry type. [line 269] [Source]

Beta - Agents - Update [Source]

• model_config body parameter gains inference_geo; note that model is whole-object replacement, so omitting inference_geo on update clears it. [line 357] [Source]

Beta - Agents - Versions - List [Source]

• Response schema now reflects the new inference_geo model config field and the BetaManagedAgentsAdvisor coordinator roster entry type. [line 273] [Source]

Beta - Deployments [Source]

• New optional budget field (BetaManagedAgentsBudgetLimit) on a deployment: a hard spend ceiling that stops new model requests once tracked cost reaches max_list_cost. [line 675] [Source]
• User message content blocks now support a new BetaManagedAgentsRedactedBlock type, a placeholder for content withheld by Anthropic model policy. [line 303] [Source]

Beta - Deployments - Archive [Source]

• Response now includes the new budget spend-ceiling field and supports BetaManagedAgentsRedactedBlock content. [line 687] [Source]

Beta - Deployments - Create [Source]

• New optional budget body parameter to set a hard spend ceiling on the deployment's sessions. [line 385] [Source]

Beta - Deployments - List [Source]

• Response items now include the new budget spend-ceiling field and support BetaManagedAgentsRedactedBlock content. [line 721] [Source]

Beta - Deployments - Pause [Source]

• Response now includes the new budget spend-ceiling field and supports BetaManagedAgentsRedactedBlock content. [line 687] [Source]

Beta - Deployments - Retrieve [Source]

• Response now includes the new budget spend-ceiling field and supports BetaManagedAgentsRedactedBlock content. [line 687] [Source]

Beta - Deployments - Unpause [Source]

• Response now includes the new budget spend-ceiling field and supports BetaManagedAgentsRedactedBlock content. [line 687] [Source]

Beta - Deployments - Update [Source]

• New optional budget body parameter to set or update a hard spend ceiling on the deployment. [line 133] [Source]

Beta - Messages - Create [Source]

• New beta feature flag mid-conversation-tool-changes-2026-07-01 added to the anthropic-beta header enum, a new opt-in beta for changing tool definitions mid-conversation. [line 101] [Source]
• claude-opus-4-1 and claude-opus-4-1-20250805 removed from the model parameter's selectable values, indicating retirement/deprecation. [line 4010] [Source]
• New documented rule for BetaThinkingBlockParam.signature: thinking blocks must be passed back to the API unmodified and in original order, or the request fails with a 400 error. [line 983] [Source]

Beta - Sessions [Source]

• New "advisor" multiagent roster entry (BetaManagedAgentsAdvisor): a coordinator's roster can now include a platform model the primary thread may consult mid-turn; at most one per roster, occupying the reserved name anthropic.advisor. [line 41] [Source]
• New session-level spend cap: BetaManagedAgentsBudgetLimit adds a budget object that stops the session from issuing new model requests once tracked list cost is reached. [line 637] [Source]
• Model config gains inference_geo: sets the geographic region for inference, falling through to the workspace's default_inference_geo when unset. [line 341] [Source]
• Session/thread usage objects gain active_seconds, list_cost, and server_tool_use for cost and runtime tracking. [line 2091] [Source]
• New periodic BetaManagedAgentsSessionUsageEvent (type: "session.usage") streams point-in-time usage/cost snapshots. [line 4587] [Source]
• Removed the previously documented 1-4096 character limit on custom tool description. [line 601] [Source]

Beta - Sessions - Archive [Source]

• Archived session response now returns budget and expanded usage fields (active_seconds, list_cost, server_tool_use), plus advisor roster entries and inference_geo. [line 939] [Source]

Beta - Sessions - Create [Source]

• New request-level budget parameter lets callers cap session spend at creation time. [line 673] [Source]
• initial_events user-message content blocks now accept a BetaManagedAgentsRedactedBlock as a placeholder for content withheld by Anthropic's model policy. [line 865] [Source]
• Model config parameter gains inference_geo, and the created-session response includes budget and expanded usage. [line 383] [Source]

Beta - Sessions - Events [Source]

• New redacted content block type added to agent-message, user-message, and other message content unions across the event schema. [line 349] [Source]
• New stop reason BetaManagedAgentsSessionBudgetReached (type: "budget_reached"): the agent stops when the session's tracked list cost hits its budget, or usage includes a model with no list price the budget can't measure. [line 3173] [Source]
• New periodic session.usage event streams point-in-time usage snapshots (tokens, active_seconds, list_cost, server_tool_use, budget) mid-session. [line 7119] [Source]

Beta - Sessions - Events - List [Source]

• Listed events now support the redacted content block, the budget_reached stop reason, and the new session.usage periodic event alongside the existing event types. [line 318] [Source]

Beta - Sessions - Events - Send [Source]

• User-message content sent to a session can now include a redacted content block placeholder in addition to text/image/document blocks. [line 272] [Source]

Beta - Sessions - Events - Stream [Source]

• Streamed events now include the redacted content-block placeholder, the new budget_reached stop reason, and the periodic session.usage event type. [line 295] [Source]

Beta - Sessions - List [Source]

• Each listed session now includes an advisor-capable roster, inference_geo on model configs, a budget field, and expanded usage. [line 759] [Source]

Beta - Sessions - Retrieve [Source]

• Retrieved session object now includes budget, advisor roster entries, inference_geo on model configs, and expanded usage. [line 689] [Source]

Beta - Sessions - Update [Source]

• New request-level budget parameter lets callers set or change a session's spend ceiling. [line 349] [Source]

Beta - Sessions - Threads [Source]

• A session thread's agent field can now resolve to a platform "advisor" entry, not just a full agent definition. [line 39] [Source]
• Thread-level usage gains active_seconds, list_cost, and server_tool_use, mirroring the session-level usage expansion. [line 845] [Source]

Beta - Sessions - Threads - Archive [Source]

• Archived thread's agent can now be a platform advisor entry, and thread usage gains active_seconds, list_cost, and server_tool_use. [line 121] [Source]

Beta - Sessions - Threads - Events - List [Source]

• Thread event history now supports the redacted content block placeholder, and SessionUpdatedEvent surfaces advisor roster entries plus the new periodic session.usage event. [line 293] [Source]

Beta - Sessions - Threads - Events - Stream [Source]

• Streamed thread events now support the redacted content block placeholder, and the new periodic session.usage event is part of the stream's event union. [line 297] [Source]

Beta - Sessions - Threads - List [Source]

• Listed threads' agent field can now resolve to a platform advisor entry, and thread usage gains active_seconds, list_cost, and server_tool_use. [line 129] [Source]

Beta - Sessions - Threads - Retrieve [Source]

• Retrieved thread's agent field can now resolve to a platform advisor entry, and thread usage gains active_seconds, list_cost, and server_tool_use. [line 121] [Source]

Beta - Skills - List [Source]

• Corrected pagination guidance: next_page should now be passed to the page query parameter (previously incorrectly documented as page_token). [line 201] [Source]

Beta - Webhooks [Source]

• New session.budget_reached webhook event, fired when a session's spend budget limit is reached. [line 36] [Source]

Build With Claude - Cache Diagnostics [Source]

• Clarifies the "same workspace" requirement for previous_message_id cache diagnostics by explaining you can verify this via the anthropic-workspace-id response header. [line 211] [Source]

Build With Claude - Claude on Amazon Bedrock (Legacy) [Source]

• New "Mid-conversation system messages on Bedrock" section: available via the InvokeModel API for Claude Fable 5 and Claude Opus 4.8 (not Sonnet 5, not ARN-versioned models), requiring no beta header. [line 189] [Source]

Build With Claude - Prompt Caching [Source]

• Corrects the minimum cacheable token count note for Claude Haiku 3.5: it is now retired except on both Amazon Bedrock and Google Cloud (previously said only Google Cloud). [line 250] [Source]

Completions - Create [Source]

• New beta feature flag mid-conversation-tool-changes-2026-07-01 added to the anthropic-beta header enum for the legacy Text Completions endpoint. [line 101] [Source]
• claude-opus-4-1 and claude-opus-4-1-20250805 removed from the model parameter's selectable values. [line 127] [Source]

Errors [Source]

• Expanded Request ID guidance: the Ruby SDK now exposes the request ID via middleware, and response-header accessors can read anthropic-organization-id and the new anthropic-workspace-id. [line 73] [Source]

Manage Claude - CMEK [Source]

• New bullet documents that, under CMEK, Compliance API local session transcripts for Cowork and Claude Code currently return no message content. [line 88] [Source]

Manage Claude - Compliance Content Data [Source]

• Page retitled from "Retrieve and delete chats, files, and projects" to "Retrieve and delete chats, files, projects, and sessions," reflecting that session transcripts are now covered by this endpoint group. [line 1] [Source]

Manage Claude - Workspaces [Source]

• The Default Workspace now has a wrkspc_-prefixed ID like any other workspace (previously documented as having no ID), retrievable via Get Workspace and the anthropic-workspace-id response header. [line 17] [Source]
• New section "Identify the workspace behind an API response" documents the anthropic-workspace-id response header, including how to use it to match usage/cost reports. [line 187] [Source]

Managed Agents - Agent Setup [Source]

• New "Pin the inference geo" subsection documents setting inference_geo ("us" or "global") on the agent's model object, including validation against the workspace's allowed_inference_geos and the multiagent requirement that all roster members share the same pin. [line 85] [Source]
• Clarified that a per-session model override that omits effort does not inherit the agent's configured effort level, since the override replaces the model object in full. [line 25] [Source]

Managed Agents - Memory [Source]

• Narrowed the create-time-only attachment rule to file resources specifically, reflecting that repository resources support other post-creation operations such as authorization token rotation. [line 52] [Source]

Managed Agents - Self-Hosted Sandboxes [Source]

• Required ant worker version bumped from 1.21.0 to 1.22.1 in both the Linux binary install and Docker build instructions. [line 100] [Source]
• Custom tool description requirement loosened from a specified 1-4,096 character range to simply "non-empty." [line 463] [Source]

Managed Agents - Sessions [Source]

• New "Pin the inference geo for a session" subsection with a full example showing how a per-session model override sets or clears the agent's inference_geo pin. [line 125] [Source]
• Clarified that a per-session model override drops the agent's own effort entirely, so such a session always runs at the model's default effort level. [line 97] [Source]

Managed Agents - Skills [Source]

• Documents that the mount_path field on a mounted github_repository resource is optional and defaults to /workspace/<repo-name>. [line 96] [Source]

Managed Agents - Webhooks [Source]

• Clarified session.thread_terminated semantics: it no longer fires when the parent session terminates; a coordinator-spawned child that finishes goes idle (not terminated), and the primary thread's end now surfaces only as session.status_terminated. [line 24] [Source]

Messages - Create [Source]

• claude-opus-4-1 and claude-opus-4-1-20250805 removed from the model parameter's selectable values (GA endpoint, mirroring the beta deprecation). [line 3117] [Source]
• ThinkingBlockParam.signature: new documented rule that thinking blocks must be passed back unmodified and in original order, otherwise the API returns a 400 error. [line 869] [Source]

Models - List [Source]

• New beta feature flag mid-conversation-tool-changes-2026-07-01 added to the supported anthropic-beta header values. [line 121] [Source]

Models - Retrieve [Source]

• New beta feature flag mid-conversation-tool-changes-2026-07-01 added to the supported anthropic-beta header values. [line 105] [Source]

Rate Limits [Source]

• New guidance recommending caching of repeated content to boost effective throughput without raising rate limits, with a pointer to monitor cache hit rate on the Usage page. [line 74] [Source]
• Documents that the anthropic-workspace-id response header can be used to identify which Workspace a request's rate limit counted against. [line 187] [Source]

Release Notes - Overview [Source]

• Adds a new August 11, 2026 entry documenting the new anthropic-workspace-id response header. [line 12] [Source]
• Clarifies that the Claude Opus 4.1 and other prior model retirements only cause errors "on the Claude API," implying continued availability on Bedrock/Google Cloud. [lines 28-283] [Source]

SDKs - Java [Source]

• Bumps the recommended Java SDK dependency version from 2.52.0 to 2.53.0. [line 14] [Source]

SDKs - Ruby [Source]

• Removes the caveat that the Ruby SDK is in "initial development" with a major version of 0 and that APIs may change at any time, signaling the SDK's versioning has stabilized. [line 421] [Source]