Anthropic · Claude
New in this edition · 37
- APIClaude API
Fable 5.2 thinking blocks can only be used by the account that generated them
Thinking blocks generated by Fable 5.2 and Mythos 5.2 work only with the account that generated them or a linked account. If another account submits one, the API drops the block; with the beta header, this may appear in input_transformations as organization_binding_mismatch.
Usage details
Who can use it: Claude API and Google Cloud Google-managed model access; integrations that share transcripts across accounts.
How to use it: If you transfer conversations across accounts, don't share thinking blocks, or handle the input_transformations field on return using the thinking-binding-controls-2026-08-01 beta header.
Tip: When storing transcripts, save the model, account, and organization IDs alongside the thinking blocks.
✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Claude Fable 5.2 ve Claude Mythos 5.2 migration guideSource: Claude Fable 5.2 yenilikleri - AvailabilityClaude
Fable 5.2 is now available on all paid plans; Enterprise admins must enable it
Fable models are now available on the Pro, Max, Ultra, Team, and Enterprise paid plans; the Free plan is excluded. On Enterprise, the model remains disabled until an admin enables Fable 5.2. Seven days after launch, Fable’s active role and defaults will be migrated to Fable 5.2.
Usage details
Who can use it: Claude web, Mobile, Desktop, Cowork, Code, Design, Microsoft 365, and Claude Tag; paid plans. Enterprise requires model access to be granted by an organization admin.
How to use it: Select Fable 5.2 from the model picker in Web, Desktop, or Mobile. An Enterprise admin must enable Fable 5.2 under model access in Organization settings.
Tip: Check in advance whether Fable usage for Pro and Team standard seat users may draw from usage credits.
✓ Verified by official sourceModels: Claude Fable 5.2Source: Claude Fable modelleri planınızdaSource: Claude Fable 5.2 ve Mythos 5.2 duyurusu - APIClaude API
Fable 5.2 and Mythos 5.2 require adaptive thinking
Thinking always runs adaptively in these models; sending budget_tokens with thinking.type set to disabled or enabled returns a 400 error. Forcing tool_choice to any or to a specific tool is also unsupported; use auto or none.
Usage details
Who can use it: Claude API and compatible platforms that offer these models; API integrations.
How to use it: Remove the thinking field from the request or send {type: "adaptive"}; control depth with effort. Use auto or none for tool_choice.
Tip: Before going to production, test legacy manual thinking and forced tool-call flows separately with the new model.
✓ Verified by official sourceModels: Claude Fable 5.2, Claude Mythos 5.2Source: Claude Fable 5.2 ve Claude Mythos 5.2 migration guideSource: Claude Fable 5.2 yenilikleri - CapabilityClaude Code
If memory runs out while reading a transcript, new messages and SDK sessions are preserved
A bug that caused new messages to be lost and claude -p, SDK, or VS Code sessions to hang when memory ran out while reading a transcript has been fixed. The session now handles the error more reliably.
Usage details
Who can use it: Claude Code 2.1.321 and later; CLI, SDK, and VS Code sessions using long transcripts.
How to use it: Update the version; continue in the affected session while keeping a transcript backup.
Tip: Regularly splitting very long sessions with a new summary or a separate session reduces memory pressure.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
AskUserQuestion responses can no longer set the hidden afkTimeoutMs field
In claude -p and SDK sessions, attempts by a host to change the timeout by adding the undocumented afkTimeoutMs field to a permission response are now rejected. Other AskUserQuestion flows remain unchanged.
Usage details
Who can use it: Claude Code CLI and Agent SDK; AskUserQuestion permission responses provided by a host application.
How to use it: Remove the afkTimeoutMs field from your host integration; continue using supported AskUserQuestion fields.
Tip: Validate SDK schemas during upgrades and avoid depending on undocumented fields.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - SettingClaude Code
Auto default now saves the correct permission selection in /config
When a new session starts in Auto mode, /config no longer shows Manual. A selection made with Space on Auto (default) is saved as Don't Ask; the previous incorrect Plan setting is no longer saved.
Usage details
Who can use it: Claude Code 2.1.321 and later; terminal sessions.
How to use it: Update Claude Code to 2.1.321 or later; open /config and select a permission mode.
Tip: Before changing the session's starting permission policy, check that your selection in /config was actually saved.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
Failed NotebookEdit no longer reports success to hooks or telemetry
If a notebook changes while the permission prompt is open and the edit fails, NotebookEdit no longer sends a success signal to hooks or telemetry.
Usage details
Who can use it: Claude Code 2.1.321 and later; hooks and OpenTelemetry pipelines using NotebookEdit.
How to use it: Update the version; evaluate the NotebookEdit result in the hook along with its success code.
Tip: In automations, check the success field of the edit result, not just whether the tool call completed.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
Ask reasons and output matching fixed for multiple PreToolUse hooks
In applications embedding Claude Code, one hook's ask reason no longer gets mixed up with another hook's reason; the previous ask response is also no longer lost when defer completes. Fixed an issue where a large hook output pointed to the wrong hook's file.
Usage details
Who can use it: Claude Code 2.1.321 and later; embedded applications and PreToolUse hooks.
How to use it: Upgrade to 2.1.321 without changing your hooks; retest results if multiple hooks inspect the same tool call.
Tip: Logging the hook name and call ID together makes debugging easier.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
Memory usage is lower in long sessions that read single lines from large files
Fixed high memory usage caused by repeatedly reading single lines from large files in long sessions. This reduces the risk of long-running sessions slowing down or closing due to memory pressure.
Usage details
Who can use it: Claude Code 2.1.321 and later; long-running terminal, SDK, and desktop sessions.
How to use it: Install the latest version; no need to migrate existing sessions separately.
Tip: For very large files, request the necessary range in a single Read call instead of reading single lines in a loop.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
The claude agents command now correctly interprets flags written before it
Options written before claude agents—--permission-mode, --dangerously-skip-permissions, --model, --effort, or --agent—are no longer ignored when used with another flag such as --verbose.
Usage details
Who can use it: Claude Code CLI 2.1.321 and later; command-line and script usage.
How to use it: Run claude --model <model> agents or a similar existing command sequence with 2.1.321 or later.
Tip: In CI scripts, use one consistent format rather than mixing options before and after the subcommand.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
In Cloud sessions, pkill, pgrep, and pidwait no longer match Claude's main processes
In Cloud sessions, pkill, pgrep, and pidwait no longer accidentally match the session's own parent processes. This reduces the risk of commands targeting Claude's own runtime infrastructure matching the wrong processes.
Usage details
Who can use it: Claude Code Cloud sessions; 2.1.321 and later.
How to use it: Start the Cloud session with version 2.1.321 or later; no additional settings are needed for the commands.
Tip: Even so, use a narrow name or PID filter in bulk process cleanup commands.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude API
High-risk cyber and biology prompts for Fable 5.2 can be routed to Opus
When Fable 5.2's safety classifiers block high-risk cyber or biology requests, Claude Opus 4.8 or Claude Opus 5 can be offered as a suitable fallback. This separates the safety block from the continuation of technical work.
Usage details
Who can use it: Claude Fable 5.2; flows where the relevant safety classifiers are applied in Claude API, Claude Code, Claude.ai, and Cowork.
How to use it: In the API, handle refusal and fallback behavior using the stop_reason/refusal fields; use fallbacks: "default" for supported server-side fallback.
Tip: Label fallback output separately from the primary model's output and log the safety category.
✓ Verified by official sourceModels: Claude Fable 5.2Source: Claude Fable 5.2 ve Mythos 5.2 duyurusuSource: Claude Fable 5.2 yenilikleri - Pricing / PlanClaude Desktop
Desktop spend limit warnings no longer disappear when the Gateway policy has no code rule
Fixed an issue that prevented Claude Desktop from showing spend limit warnings when the Claude apps gateway policy contained no code rule at all. Spend limit notifications are visible again behind a compatible Gateway.
Usage details
Who can use it: Organizations using Claude Desktop and Claude apps gateway; spend limit warnings.
How to use it: Upgrade Gateway and Desktop to 2.1.321 or later; do not change the existing spend limit setting.
Tip: Test whether users see limit warnings with a real limit scenario behind Gateway.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude apps gateway
Gateway warns about new desktop settings instead of rejecting them outright
When a desktop setting is newer than the version supported by Gateway, such as the oauth.omitResource field in a managed MCP server, Gateway no longer refuses to start. It warns about the unknown new field and continues serving.
Usage details
Who can use it: Claude apps gateway 2.1.321 and later; organizations using new Claude Desktop managed settings.
How to use it: Update Gateway to 2.1.321; review the log warning and plan to move to a Gateway version that supports the field.
Tip: When using a new field in managed settings, upgrade the Gateway and Desktop versions together.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
When the first model is blocked before responding, the fallback model no longer gives an incorrect refusal
Fixed an issue where the fallback model would sometimes refuse the request or refer to an interrupted response when the first model was blocked before responding. The fallback now handles the new request with the correct context.
Usage details
Who can use it: Claude Code 2.1.321 and later; sessions using model fallback.
How to use it: Use an up-to-date version of Claude Code; no fallback configuration or model selection changes are needed.
Tip: In fallback tests, separately measure the scenario where the first model is blocked before responding.
- AvailabilityClaude Code
If macOS MDM policy loads slowly, it is retried once at startup
If a valid MDM profile loads slowly, Claude Code now retries managed policy settings once during macOS startup. Timeouts are also reported explicitly as timeouts instead of general read errors.
Usage details
Who can use it: Claude Code macOS; devices managed with MDM.
How to use it: Update Claude Code; allow the MDM profile to be read on the device and restart the session.
Tip: For enterprise deployments, schedule MDM profile application before Claude Code starts.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - Plugin / IntegrationClaude Code
Old settings and secrets from managed marketplace takeovers can be restored
When a managed extraKnownMarketplaces source takes over another source, old entries are set aside for the duration of cleanupPeriodDays. Settings and secrets can be restored with claude plugin restore-set-aside.
Usage details
Who can use it: Organizations using the Claude apps gateway or managed marketplace policies; Claude Code 2.1.321 and later.
How to use it: After upgrading, review the entries that were taken over; use the claude plugin restore-set-aside command if needed.
Tip: Before migrating marketplaces, export existing plugin options and secrets to create a rollback plan.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - Plugin / IntegrationClaude Code
Claude in Chrome processes no longer linger after closing behind a proxy
Fixed an issue where Claude in Chrome and computer use server processes continued running after the parent process closed when a proxy connection stalled. Orphaned browser processes are now cleaned up more reliably when a session ends.
Usage details
Who can use it: Claude Code, Claude in Chrome, and computer use; sessions behind a proxy.
How to use it: Upgrade Claude Code to 2.1.321 or later; you may need to manually terminate existing orphaned processes once.
Tip: If you use long keep-alive periods on a proxy, monitor session shutdown logs.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - AgentClaude Code
Pending background agent reports after Stop no longer overload a single request
Fixed an issue where too many pending background agent reports were sent in a single request after Claude was stopped. The main conversation now receives a more controlled amount of report content.
Usage details
Who can use it: Claude Code 2.1.321 and later; terminal and desktop sessions using background agents.
How to use it: Use the latest version; retest workflows that stop and restart many agents at once.
Tip: Splitting agent tasks into smaller groups also reduces context usage in the main session.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
Prompt suggestions no longer make extra API calls for responses ending in a tool call
Fixed an issue where prompt suggestions sent unnecessary additional API requests when the model responded with a tool call. This reduces unnecessary latency and usage.
Usage details
Who can use it: Claude Code 2.1.321 and later; sessions with prompt suggestions enabled.
How to use it: Update the version; no additional setting or beta header is required.
Tip: Compare the number of requests generated by prompt suggestions before and after updating when tracking API costs.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
Windows watchPaths hooks no longer stop on lowercase drive paths
A bug that caused FileChanged hooks to stop after the first shell command when the working directory’s drive letter was lowercase on Windows has been fixed. File changes monitored with watchPaths now trigger throughout the session.
Usage details
Who can use it: Claude Code Windows sessions; projects using watchPaths.
How to use it: Update Claude Code to 2.1.321 or later; continue using watchPaths without changing your settings.
Tip: Using a consistent drive-letter format in Windows configurations can still help reduce path-comparison issues.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
If TLS disconnects mid-response, the partial response is preserved or the request is retried
When a connection drops during a response and causes an EPROTO TLS error, Claude Code now preserves the partial response or retries the request. This restores the resilient behavior of previous versions.
Usage details
Who can use it: Claude Code 2.1.321 and later; API, proxy, or gateway environments with unstable connections.
How to use it: Update the version; in environments that use a proxy, restart the regular Claude Code session.
Tip: For critical, long-running tasks, keep transcript backups enabled so you don’t lose partial response text.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - AvailabilityClaude Code
Unresponsive TLS networks no longer leave sign-in and token operations waiting indefinitely
Fixed an issue where sign-in, token refresh, and logout operations could wait indefinitely if a network connection was accepted but no TLS handshake response arrived.
Usage details
Who can use it: Claude Code 2.1.321 and later; environments using a proxy, VPN, or TLS inspection.
How to use it: Install the latest version; if the connection issue persists, check the proxy’s TLS handshake response and timeout settings.
Tip: Wrap authentication commands in a timeout in CI so network failures do not leave jobs running indefinitely.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - AgentClaude Code
Subagent progress summaries no longer run unnecessarily to completion when the model returns a tool call
Subagent progress summaries no longer keep running unnecessarily until completion after the model responds with a tool call. This reduces wait time and unnecessary model calls.
Usage details
Who can use it: Claude Code 2.1.321 and later; background subagent progress summaries.
How to use it: Update the version; start subagents as usual.
Tip: When running many subagents, monitor progress notifications’ call counts in telemetry.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - FeatureClaude Code
Inter and Space Grotesk added to the recommended defaults in the Artifact design guide
Artifact design guidelines now consider the Inter and Space Grotesk fonts suitable default options. This affects Claude’s typography choices when generating new Artifacts.
Usage details
Who can use it: Claude Artifact generation; updated design guide with Claude Code 2.1.321 and later.
How to use it: When generating an Artifact, use the default design guidance unless you want a specific font; if you need a particular brand font, state it explicitly in the prompt.
Tip: For accessibility, request a readable body font instead of a decorative font.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - FeatureClaude Code
Guidance on signing in and copying Artifacts from another organization is clearer
When an Artifact belongs to another organization, Claude now recommends /login only if the Artifact is in one of your organizations. It also explains more clearly where to find a copy after an account transfer.
Usage details
Who can use it: Multi-organization users working with Claude Code and Artifact; 2.1.321 and later.
How to use it: Check the organization information in the Artifact message; sign in to the correct organization account, then copy the Artifact if needed.
Tip: Export important Artifact links and copies before switching organizations.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
Skill and command names containing Bamum characters that appear blank in some fonts are rejected
Skill and command names containing a Bamum letter that some fonts render as blank are no longer accepted. This prevents the use of names that are invisible or indistinguishable in menus.
Usage details
Who can use it: Claude Code 2.1.321 and later; skill and slash command authors.
How to use it: Rename the skill or command using Latin letters and safe separators, then reload the plugin.
Tip: Limit names in shared plugins to visible Unicode characters.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - FeatureClaude Code
Sorting guidance for large Artifact databases now recommends reading page by page
When an Artifact database collection is too large to sort, the Artifact tool no longer suggests adding filters. Claude is instructed to read and sort the collection page by page.
Usage details
Who can use it: Claude Code and the Artifact tool; sessions using large database collections.
How to use it: Use the Artifact database in the latest version of Claude Code; follow Claude’s pagination guidance for large collections.
Tip: For large collections, limit the page size and required fields before sorting.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
claude -p now reports the duration of background commands accurately
claude -p no longer incorrectly reports that long-running background commands stopped after 10 minutes took 30 minutes. The duration reported to Claude now matches the actual behavior.
Usage details
Who can use it: Claude Code 2.1.321 and later; claude -p and automation runs.
How to use it: Update Claude Code; no additional setting needs to be enabled or disabled.
Tip: In automation prompts, use the actual status reported by the command result rather than assuming a fixed duration.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - FeatureClaude Code
The dataviz skill recommends a separate meter row for each ratio
The bundled dataviz skill now recommends a separate meter row for each ratio when multiple ratios have their own limits. This makes it possible to see each limit separately instead of using a single overall meter.
Usage details
Who can use it: Claude Code 2.1.321 and later; chart generation using the bundled dataviz skill.
How to use it: Specify the ratios and limits in your dataviz request; the updated skill will recommend a separate row for each one.
Tip: In dashboard designs, include the ratio's unit and limit in the meter heading.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
The command panel and prompt reappear after a dialog closes
Fixed an issue where the command panel disappeared after a dialog closed and the prompt remained hidden until a redraw. The panel and input field now return as expected.
Usage details
Who can use it: Claude Code terminal interface 2.1.321 and later.
How to use it: Update Claude Code; if the issue occurs, reopen with the latest version rather than waiting for a redraw in the same session.
Tip: If you see a temporary blank screen in the terminal, check the panel state first instead of clearing the entire session with Ctrl+L.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - CapabilityClaude Code
Install scripts and trusted TLS certificate loading are faster
claude install and install scripts now complete about 2-4 seconds faster. Trusted TLS certificates also load faster at startup.
Usage details
Who can use it: Claude Code 2.1.321 and later; installation and every session startup.
How to use it: Rerun claude install or the existing install script with the latest version.
Tip: When comparing enterprise CI times, measure first runs and cached runs separately.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
Unnecessary SEP-2352 warning removed when logging in to MCP again
When logging in again to an MCP server saved by Claude Code, claude mcp login no longer prints a SEP-2352 issuer warning every time.
Usage details
Who can use it: Claude Code 2.1.321 and later; users logging in again to saved MCP servers.
How to use it: Run claude mcp login <server> as usual; no additional setting is required.
Tip: If you still see a warning after logging in again to an OAuth server, check the version first instead of deleting and re-adding the server.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
Non-interactive telemetry hook counts no longer double-count built-in hooks
After Bash and PowerShell calls, two built-in hooks are no longer added to the num_hooks and num_success fields in OpenTelemetry hook events for non-interactive sessions.
Usage details
Who can use it: Claude Code non-interactive sessions; OpenTelemetry hook metrics.
How to use it: Update to 2.1.321; use the new counter baseline in your telemetry consumer.
Tip: Add a version breakdown when comparing counters from old and new versions on the same chart.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - APIClaude Code
OpenTelemetry vcs fields are no longer missing in previously launched Desktop processes
OpenTelemetry repository attributes are no longer missing from Claude Desktop sessions opened from a previously launched process. vcs.* fields are now correctly included in the telemetry pipeline.
Usage details
Who can use it: Enterprise monitoring setups using the Claude Desktop Code tab and OpenTelemetry.
How to use it: Update Claude Code; open a new session without reconfiguring your exporter and check the vcs.* fields.
Tip: In dashboards, read repository identity from vcs.* fields as well as process-start labels.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - SettingClaude Code
Remote Control’s initial prompt accepts a y response without a trailing newline
If terminal input in the initial claude remote-control prompt ends with y and does not include Enter, Claude Code no longer silently ignores it. The response is treated as confirmation.
Usage details
Who can use it: Claude Code 2.1.321 and later; Remote Control launched with scripts, pipes, and batch jobs.
How to use it: Update the claude remote-control command; if you send y to the initial prompt through a pipe, adding a separate Enter is no longer required.
Tip: For automation, printf 'y\n' is still more portable across different terminal environments.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes - Plugin / IntegrationClaude Code
Built-in plugin type files are no longer written accidentally to the builtin folder
Fixed an issue where type files generated for Claude Code's built-in plugins created an unwanted builtin folder in the project when using --plugin-dir . or --plugin-dir ./builtin.
Usage details
Who can use it: Claude Code 2.1.321 and later; plugin developers using --plugin-dir.
How to use it: Update the version and restart the plugin with claude --plugin-dir <folder>.
Tip: Before adding generated files to version control, check the working tree for an unexpected builtin folder.
✓ Verified by official sourceSource: Claude Code v2.1.321 release notes