Agent SDK billing change (June 15, 2026): claude -p and Agent SDK usage on subscription plans will draw from a new separate monthly credit, distinct from interactive usage limits — announced across multiple docs
claude agents gains session defaults and config flags: new --permission-mode, --model, and --effort flags set defaults for all dispatched sessions; new section documents how to pass settings, plugins, and MCP servers to the agent view
Output styles page overhauled: restructured with a numbered creation workflow, clearer guidance on when to use keep-coding-instructions, and a new comparison table against CLAUDE.md, --append-system-prompt, agents, and skills
Model lifecycle updates: Claude Sonnet 3.7 marked as fully retired on Bedrock (April 28) and Vertex AI (May 11); explicit retirement dates added for Opus 4, Sonnet 4, and Haiku 3.5 across platforms; removed from tool/pricing tables throughout
Context window overflow behavior clarified: Claude 4.5+ now accepts requests exceeding the context window and stops with stop_reason: "model_context_window_exceeded" instead of a validation error
Added terminalSequence field to hook JSON output so hooks can emit desktop notifications, window titles, and terminal bells without needing a controlling terminal
Added CLAUDE_CODE_PLUGIN_PREFER_HTTPS environment variable to clone GitHub plugin sources over HTTPS instead of SSH, for environments without a GitHub SSH key
Added ANTHROPIC_WORKSPACE_ID environment variable for workload identity federation — scopes the minted token to a specific workspace when the federation rule covers more than one
Added claude agents --cwd <path> to scope the session list to a specific directory
/feedback can now include recent sessions (last 24 hours or 7 days) for issues that span more than the current session
Rewind menu: added "Summarize up to here" to compress earlier context while keeping recent turns intact
Auto mode permission dialog now explains when a permissions.ask rule caused the prompt
Restored the "view diff in your IDE" option on file-edit permission prompts when an IDE is connected
Existing feature improvements
Background agents launched via /bg or ←← now preserve the current permission mode instead of reverting to default
claude agents: agents that finish work but leave a background shell running now move to Completed instead of staying as Working
Spinner warms to amber after 10 seconds of thinking to signal Claude is still working
Improved plugin menu navigation: →/Tab switch tabs, ↑ moves to the tab strip, and tab headers and search box are clickable in fullscreen mode
Background side-queries on Bedrock/Vertex/Foundry/gateway now fall back to the main-loop model when no ANTHROPIC_SMALL_FAST_MODEL is set, instead of sending an unavailable Haiku model ID
claude agents launching no longer fails when the pre-warmed background worker is unhealthy — now falls back to a fresh launch
Empty idle background sessions left over from ← are automatically retired by the daemon after 5 minutes
Major bug fixes
Fixed /model in one session silently changing the autocompact threshold in other concurrent sessions
Fixed hooks receiving a non-existent transcript_path after EnterWorktree switches the working directory
Fixed markdown tables with cell wrapping falling back to the vertical key-value layout instead of rendering as a bordered grid (regression in 2.1.136)
Fixed Ctrl+C not interrupting a running turn while in vim INSERT/VISUAL mode
Fixed alternative chat:submit keybindings (e.g. meta+enter, ctrl+enter) not working when enter is rebound to chat:newline
Fixed prompt suggestions being silently disabled when an output style was configured
Fixed /tui silently dropping running background shells and subagents — now refuses and asks to wait for them to finish
Fixed Bedrock awsCredentialExport being skipped when ambient AWS credentials resolve, fixing auth for cross-account access
Fixed Remote Control MCP connectors all failing with 401 when the worker session token rotated mid-session
Fixed SDK "Claude Code native binary not found" on Linux when both glibc and musl platform packages are installed
Fixed MCP HTTP/SSE servers returning 403 on connect showing as "failed" instead of "needs auth"
Fixed plugin details pane showing 0 MCP servers for plugins that declare them via .mcp.json
Fixed Web Search status showing "Did 0 searches" when searches returned errors
Added a note that several Claude Code features (web, Routines, Code Review, Remote Control, Chrome extension) require a Claude.ai account and are not available through Console API keys or cloud-provider credentials alone — relevant for teams deploying via Bedrock, Vertex, or Foundry [line 18] [Source]
Clarified that a custom output style replaces claude_code preset's software engineering instructions by default; set keep-coding-instructions: true in the frontmatter to keep them and layer your instructions on top [lines 30-34] [Source]
Updated the code-review example to include keep-coding-instructions: true in the frontmatter [line 42] [Source]
Updated comparison table: output styles column now reads "Replace or extend default" to reflect the new keep-coding-instructions option [line 51] [Source]
Added notice: starting June 15, 2026, Agent SDK and claude -p usage on subscription plans will draw from a new monthly Agent SDK credit, separate from interactive usage limits [line 9] [Source]
"Permission mode and settings" section renamed to "Permission mode, model, and effort" and expanded: claude agents now accepts --permission-mode, --model, and --effort flags that set defaults for every session dispatched from that agent view instance; active defaults appear in the footer [lines 108-121] [Source]
New "Settings, plugins, and MCP servers" section documents how to pass --settings, --add-dir, --plugin-dir, --mcp-config, and --strict-mcp-config to claude agents — these apply to agent view itself and to every session dispatched from it [lines 123-141] [Source]
Clarified that backgrounding a session also does not transfer running monitors (in addition to subagents and background commands) [line 91] [Source]
Removed the "through May 5, 2026" expiration date from the /ultrareview free runs description — the 3 free runs are now described as a standing one-time allotment with no expiry date [line 181] [Source]
Clarified that disableAllHooks: true only disables user-configured hooks — hooks configured in managed settings continue to run unless disableAllHooks is also set there [line 226] [Source]
Updated the CLAUDE.md locations table to show explicit load order from broadest to most specific scope; corrected the row ordering so "User instructions" appears before "Project instructions", matching actual load order [lines 252-260] [Source]
Page substantially rewritten and restructured: custom output style creation is now presented as a 3-step numbered guide (create file → add frontmatter/instructions → switch to style) [lines 346-393] [Source]
Clarified that custom styles exclude Claude Code's built-in software engineering instructions unless keep-coding-instructions: true is set in the frontmatter [line 273] [Source]
Comparisons section replaced with a concise table comparing output styles, CLAUDE.md, --append-system-prompt, agents, and skills [lines 447-454] [Source]
Added "Related resources" section with links to settings, permission modes, plugins, and debug config [lines 456-461] [Source]
Clarified deny rule precedence: a user-level deny blocks a project-level allow (and vice versa) because deny rules from any scope are evaluated before allow rules [line 473] [Source]
Clarified that both force-enabled and force-disabled plugins in managed settings cannot be overridden by --plugin-dir (previously only force-enabled was mentioned) [line 486] [Source]
Clarified how scopes interact: scalar settings from higher-priority scopes override lower ones; array settings (like permission rules) concatenate and deduplicate across scopes rather than replace [lines 499-527] [Source]
Updated examples to distinguish scalar vs. array setting behavior and avoid using permission rules as the example for simple override [lines 508-517] [Source]
Removed the "expire on May 5, 2026" language from the free runs description — the 3 free runs are now described as a permanent one-time allotment that does not refresh or expire [lines 540-545] [Source]
Removed references to Claude Sonnet 3.7 migration; updated summarization description to remove the "easy migration from Claude Sonnet 3.7" framing [lines 168-175] [Source]
Removed older models (Claude Sonnet 3.7, Haiku 3.5, Opus 3, Haiku 3) from the Batches API pricing table; added deprecation/retirement labels to Claude Opus 4, Sonnet 4, and Haiku 3.5 [lines 187-198] [Source]
Added a note explaining that lifecycle terms and dates in the Bedrock table are set independently by AWS and may differ from the Anthropic-operated schedule [line 210] [Source]
Updated model table: Claude Sonnet 3.7 now marked as "Retired April 28, 2026" with all regions removed; Claude Opus 4 retiring May 31 2026; Claude Sonnet 4 retiring October 14 2026; Claude Haiku 3.5 retiring June 19 2026 [lines 219-227] [Source]
Added a note explaining that lifecycle terms and dates in the Vertex AI table are set independently by Google Cloud [line 239] [Source]
Updated model table: Claude Sonnet 3.7 marked as "Retired May 11, 2026"; Claude Opus 4 and Sonnet 4 retiring September 14 2026; Claude Haiku 3.5 retiring July 5 2026 [lines 247-257] [Source]
Updated beta header documentation: removed references to Sonnet 3.7 from the beta version table and added Claude Haiku 4.5 and Opus 4 ([deprecated]) to the computer-use-2025-01-24 list [lines 27-28] [Source]
Simplified the sampling loop code example: removed thinking_budget parameter, the text_editor_type variable, and the optional thinking block — uses text_editor_20250728 unconditionally [lines 47-89] [Source]
Enhanced actions description changed from "Available in Claude 4 models and Claude Sonnet 3.7" to "Available on all models that support computer use" [line 98] [Source]
Removed Claude Sonnet 3.7 row from the tool token pricing table [line 120] [Source]
Section renamed "Context window overflow behavior" and content updated: on Claude 4.5 models and newer, requests that exceed the context window are accepted and generation stops with stop_reason: "model_context_window_exceeded"; older models still return a validation error but can opt in with the model-context-window-exceeded-2025-08-26 beta header [line 282] [Source]
Removed Claude Sonnet 3.7 from the thinking model comparison table; Opus 4.7 now shows thinking output as "Omitted by default" (matching Opus 4.6), and the table is condensed from 7 columns to 6 [lines 365-374] [Source]
Updated context window behavior: on Claude 4.5+ models, max_tokens exceeding the context window now stops with model_context_window_exceeded instead of a validation error; older behavior documented alongside [line 335] [Source]
Removed note about Claude Sonnet 3.7 returning full (non-summarized) thinking output [lines 321-322] [Source]
Simplified the supported models description: images are now described as supported on "all current Claude models" rather than "all Claude 3+ models"; PDF and code execution tool support now deferred to their respective linked pages [line 387] [Source]
Updated the tool token counts table: removed Claude Sonnet 3.7, Haiku 3.5, Opus 3, Sonnet 3, and Haiku 3; added deprecation/retirement labels to Claude Opus 4, Sonnet 4, and Haiku 3.5 [lines 132-144] [Source]
Updated pricing table: removed Claude Sonnet 3.7, Haiku 3.5, Opus 3, and Haiku 3; added deprecation/retirement labels to Claude Opus 4, Sonnet 4, and Haiku 3.5 [lines 408-419] [Source]
Updated minimum cacheable prompt lengths: Claude Sonnet 4.6 minimum lowered from 2,048 to 1,024 tokens (grouped with Sonnet 4.5, Opus 4.1, Opus 4, Sonnet 4); cache minimums section now notes platform differences and adds a separate callout for Bedrock's own prompt caching documentation [lines 428-447] [Source]
Removed Claude Sonnet 3.7 from the supported models list; updated Claude Haiku 3.5 status to "retired, except on Bedrock and Vertex AI" [lines 468-471] [Source]