Every update, on the record.
Claude changes constantly. So does this manual. Every update to a feature lands here — proof the product stays current.
- FEATURE
Compliance API session endpoints exit beta for Cowork and Claude Code
The Compliance API session endpoints are now generally available for Cowork and Claude Code sessions. Local session endpoints also add beta support for Claude Science and Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook for Claude Enterprise organizations.
Official source → - FEATURE
Claude Code 2.1.251: Hooks, Subagent Streaming, Spend Limits & 50+ Fixes
Claude Code 2.1.251 adds model-switch hook events, live streaming of foreground subagent tool calls to Remote Control clients, spend-limit visibility in /usage, and per-session prompt-cache stats in /cost, alongside more than 50 bug fixes and improvements.
Official source → - FEATURE
Claude Code 2.1.248: Restricted Mode, Cache TTL, Cross-Session Messaging & Dozens of Fixes
Version 2.1.248 adds a `--restricted` flag for locked-down execution environments, per-agent prompt-cache TTL control, cross-session messaging on Bedrock/Vertex/Foundry, and a `/usage-credits` command for Enterprise users. The release also ships over 30 bug fixes covering prompt-cache misses, agent-view session handling, MCP, hooks, and security-sensitive file uploads.
Official source → - FEATURE
Claude Code 2.1.246: Major bug fix wave, new permissions UI, and stability improvements
Version 2.1.246 of Claude Code ships two new features—an Auto mode tab in `/permissions` and a completion timestamp on the end-of-turn line—alongside dozens of bug fixes covering fullscreen rendering, plugin management, background sessions, MCP tools, themes, and more.
Official source → - FEATURE
Claude Code 2.1.243: Usage Visibility, Model Control & Dozens of Fixes
Version 2.1.243 adds loop-level usage breakdowns, a configurable model picker, prompt-cache TTL controls, org-level pricing overrides, and keyless Console sign-in, alongside more than 30 bug fixes and notable startup and memory improvements.
Official source → - FEATURE
Claude Sonnet 5 launches with 1M token context and new tokenizer
Claude Sonnet 5 (`claude-sonnet-5`) is now available at $2/$10 per MTok introductory pricing, offering a 1M token context window and 128k max output tokens. Three breaking behavior changes apply on migration: adaptive thinking is on by default, manual extended thinking is removed, and non-default sampling parameters now return a 400 error.
Official source → - FEATURE
Python SDK v1.0 released with httpx2, Python 3.10+ requirement
Anthropic released v1.0 of the Python SDK, swapping the HTTP layer from httpx to httpx2 and removing long-deprecated APIs including the legacy Text Completions API and several Messages parameters.
Official source → - FEATURE
Claude Code for Windows gains cross-session messaging (v2.1.239)
Claude Code 2.1.239 brings cross-session messaging to Windows, letting Claude Code sessions on different machines message each other via `SendMessage` and discover each other via `ListAgents` — matching the existing macOS and Linux behavior.
Official source → - FEATURE
Claude Code 2.1.239: Cost accuracy, migration tooling, and 40+ fixes
Version 2.1.239 of Claude Code adds data-residency cost transparency, a Python SDK migration command, synced plugin labeling, and resolves over 40 bugs across streaming, sessions, UI, and platform compatibility.
Official source → - FEATURE
MCP Tunnels API moved to /v1/tunnels on the Claude API (research preview)
The MCP tunnels management API has migrated from `/v1/organizations/tunnels` on the Admin API to `/v1/tunnels` on the Claude API. The previous endpoint remains available during a migration window.
Official source → - FEATURE
Admin API user-management endpoints for Claude Enterprise now GA
The Admin API endpoints for managing Claude Enterprise organization members, invites, groups, and custom roles are now generally available. The `anthropic-beta: ce-user-management-2026-07-13` beta header is no longer required for group and custom-role requests.
Official source → - FEATURE
Claude Opus 5 launched with 1M token context and thinking on by default
Claude Opus 5 (`claude-opus-5`) is now available with a 1M token context window, 128k max output tokens, and thinking enabled by default — at the same $5/$25 per MTok pricing as Claude Opus 4.8.
Official source → - FEATURE
New `agent-memory-2026-07-22` beta header changes memory listing behavior
Anthropic has introduced the `agent-memory-2026-07-22` beta header, which tightens and stabilizes how the list-memories endpoint works. On July 22, 2026, the older `managed-agents-2026-04-01` header will adopt the same behavior.
Official source → - MODEL
Fast Mode Removed for Claude Opus 4.6
Anthropic has removed fast mode support for Claude Opus 4.6. Requests using `speed: "fast"` on that model now run at standard speed and standard billing rates, with no error returned.
Official source → - MODEL
Claude Opus 5: 1M token context, thinking on by default
Claude Opus 5 launches with a 1M token context window, 128k max output tokens, and thinking enabled by default — at the same $5/$25 per MTok pricing as Claude Opus 4.8.
Official source → - FEATURE
Compliance API now returns transcripts of local Cowork and Claude Code sessions
Claude Enterprise organizations can now use the Compliance API to retrieve transcripts and metadata from Cowork and Claude Code sessions running on users' machines, via three new beta endpoints using existing credentials.
Official source → - FEATURE
Dreams (research preview) now supports Claude Opus 5
Anthropic's Dreams managed-agent framework, currently in research preview, has added support for Claude Opus 5. Check the Supported Models section of the Dreams docs for details.
Official source → - FEATURE
Claude Code 2.1.234: Auto-resume, GitLab MR badges, security hardening, and dozens of fixes
Claude Code 2.1.234 adds automatic session continuation when usage limits reset, a GitLab merge request badge in the footer, a new `selection:clear` keybinding, and a `CLAUDE_CODE_PROJECT_DIR_NAME` environment variable for custom config directories. The release also closes a Windows NT-namespace path vector across several file-access points and ships more than 25 bug fixes.
Official source → - FEATURE
Claude Code 2.1.232: Multi-session messaging, GitLab support, security hardening
Claude Code 2.1.232 ships cross-session messaging via `@` mentions and `SendMessage`, subagent forking on by default, GitLab plugin marketplace support, and a large batch of security fixes and Remote Control reliability improvements.
Official source → - FEATURE
Claude Code 2.1.228: Stability, Security & UX Fixes
Version 2.1.228 ships over a dozen bug fixes and improvements to Claude Code, covering rendering freezes, Windows Git detection, cross-session messaging, self-hosted runners, and skill security hardening.
Official source → - FEATURE
Claude Code 2.1.227: Bug fixes and slash-command menu polish
Version 2.1.227 fixes several bugs—including broken Bash commands in GitHub Actions and a mis-triggered usage-credits prompt for Max plan users—and sharpens the slash-command menu's visual design and event-loop performance.
Official source → - FEATURE
Claude Code 2.1.224: Self-Hosted Runners, Cross-Session Messaging, and Major Fixes
Claude Code 2.1.224 adds self-hosted runner support, cross-session messaging between Claude Code instances, archive-based plugin installs, and a wave of sandbox security fixes and Remote Control improvements.
Official source → - FEATURE
Claude Code 2.1.222: Isolation fixes, security patches, and 20+ bug fixes
Version 2.1.222 closes a worktree isolation gap that let subagents run destructive git commands against the main checkout, patches a hook bypass in background agent tasks, and ships roughly 20 additional bug fixes and improvements across usage reporting, proxy handling, screen readers, and more.
Official source → - FEATURE
Claude Code 2.1.221: Focus View, Credential Masking, and Major Bug Fixes
Version 2.1.221 of Claude Code adds a VSCode Focus view that collapses tool activity behind per-turn summaries, introduces credential file masking for sandboxed Linux/WSL environments, and ships a large batch of bug fixes and workflow improvements.
Official source → - FEATURE
Claude Code 2.1.218
Version 2.1.218 of Claude Code ships a broad set of fixes and improvements covering background subagent code review, accessibility enhancements for screen readers, Windows path corruption, multi-line paste handling, MCP diagnostics, Bedrock billing, and dozens of stability and correctness fixes across the editor, agent, and session layers.
Official source → - FEATURE
Claude Code 2.1.217: Emoji Autocomplete, Subagent Caps, and Dozens of Fixes
Version 2.1.217 of Claude Code adds emoji shortcode autocomplete in the prompt input and a hard cap on concurrently-running subagents, alongside a large batch of bug fixes covering memory leaks, Windows auto-update failures, session isolation, and more.
Official source → - FEATURE
Claude Code 2.1.216: Performance fixes, agent reliability, and session stability
Version 2.1.216 delivers a significant performance fix for long sessions where message normalization was slowing Claude Code to a crawl, alongside dozens of bug fixes covering agent resumption, worktree isolation, OAuth token handling, terminal rendering, and more.
Official source → - FEATURE
Claude Code 2.1.215: Manual control over /verify and /code-review
Claude Code no longer runs the /verify and /code-review skills automatically. You now invoke them explicitly by typing /verify or /code-review when you want them.
Official source → - FEATURE
Claude Code 2.1.214: Permission Hardening, EndConversation Tool & Dozens of Fixes
Version 2.1.214 closes multiple permission-check bypasses across Bash, PowerShell, and Docker, adds the EndConversation tool so Claude can exit abusive sessions, and ships a large batch of Windows, background-session, and observability fixes.
Official source → - FEATURE
Claude Code 2.1.212
Version 2.1.212 of Claude Code ships major session-management improvements — including a redesigned /fork that runs copies as true background sessions, per-session caps on web searches and subagent spawns, and automatic backgrounding of slow MCP tool calls — alongside dozens of bug fixes across worktrees, /ultrareview, hooks, Windows daemon, and agent teams.
Official source → - FEATURE
Claude Code 2.1.211: Subagent streaming, security fixes, and ~30 bug fixes
Version 2.1.211 adds a flag to include subagent text and thinking in stream-json output, patches a permission-preview character-spoofing vulnerability, and resolves roughly 30 bugs spanning auth, MCP reconnection, background agents, file uploads, and rendering performance.
Official source → - FEATURE
Claude Code 2.1.210
Claude Code 2.1.210 ships a live elapsed-time counter on long-running tool calls, a startup warning for deprecated permission-rule syntax, and more than two dozen bug fixes covering agent isolation, session stability, paste handling, and prompt-injection hardening.
Official source → - FEATURE
Claude Code 2.1.208
Claude Code 2.1.208 ships a screen reader mode, vim insert-mode remaps, a corporate process wrapper, and mouse support for fullscreen menus, alongside dozens of bug fixes covering memory leaks, background-agent reliability, streaming errors, and rendering performance.
Official source → - FEATURE
Claude Code 2.1.199
Version 2.1.199 delivers stacked slash-skill support (up to 5 at once), smarter error surfacing for SSL and streaming failures, and a large sweep of background-agent and subagent reliability fixes.
Official source → - FEATURE
Claude Code Auto Mode Now On by Default on Bedrock, Vertex AI, and Foundry
Claude Code 2.1.207 removes the `CLAUDE_CODE_ENABLE_AUTO_MODE` opt-in requirement on Bedrock, Vertex AI, and Foundry. Auto mode is now enabled by default and can be turned off via `disableAutoMode` in settings.
Official source → - FEATURE
Claude Code 2.1.207: Auto Mode GA on Bedrock/Vertex/Foundry, Major Bug Fixes
Auto mode is now enabled by default on Bedrock, Vertex AI, and Foundry (no opt-in flag required), and this release delivers over a dozen bug fixes covering terminal freezing, auto-updater behavior, agent team crashes, Remote Control reliability, and security hardening.
Official source → - FEATURE
Claude Code 2.1.206
Claude Code 2.1.206 ships directory-path suggestions for /cd, a /doctor check that recommends trimming CLAUDE.md files, smarter git-push handling in /commit-push-pr, background-agent pre-warming, and roughly two dozen bug fixes across the CLI, agents view, MCP servers, and model picker.
Official source → - FEATURE
Claude Code 2.1.205
Claude Code 2.1.205 ships a hardened auto mode, a rebuilt /doctor checkup command, improved agent-view readability, and a large batch of bug fixes covering Windows crashes, background-agent status tracking, JSON schema handling, and more.
Official source → - FEATURE
Claude Code 2.1.204: Fix for SessionStart hook streaming in headless sessions
Version 2.1.204 fixes a bug where hook events failed to stream during SessionStart hooks in headless sessions, which could cause remote workers to be incorrectly idle-reaped mid-hook.
Official source → - FEATURE
Claude Code 2.1.203: Background Agent Reliability & UI Polish
Version 2.1.203 ships a large wave of background-agent bug fixes, several new visibility features (login-expiry warnings, manual-permission badge, MCP roots support), and meaningful performance and binary-size improvements.
Official source → - FEATURE
Claude Code 2.1.202: Dynamic Workflow Sizing, OTel Attributes & 15+ Fixes
Version 2.1.202 adds a configurable dynamic workflow size setting and new OpenTelemetry attributes for workflow-spawned agents, alongside a wide sweep of bug fixes covering Remote Control, voice dictation, session handling, and more.
Official source → - FEATURE
Claude Code 2.1.200: Permission Mode Rename, Background Agent Fixes
Version 2.1.200 renames the 'default' permission mode to 'Manual' across all surfaces and fixes a wide range of background agent reliability, session, and accessibility issues.
Official source → - FEATURE
Claude Code: Subagents treat parent-agent messages as normal task direction
In Claude Code 2.1.198, subagents now treat messages from the agent that launched them as ordinary task direction. Importantly, a parent agent's message is still never treated as the user's approval.
Official source → - FEATURE
Claude Code: Focus mode now tracks subagent activity and folds notifications
In Claude Code 2.1.198, focus mode has been improved so that subagents launched during a turn now appear in the turn's activity summary. Completed background task notifications are also folded into a single count instead of stacking individually.
Official source → - FEATURE
Claude Code: Background agents now auto-commit, push, and open draft PRs
In Claude Code 2.1.198, background agents launched from `claude agents` that finish code work in a worktree will now automatically commit, push, and open a draft pull request — instead of pausing to ask for input.
Official source → - FEATURE
Claude Code 2.1.198: Claude in Chrome GA, background agent upgrades
Claude Code 2.1.198 brings Claude in Chrome to general availability, adds automatic PR creation when background agents finish code work, introduces a /dataviz skill, and ships a large round of bug fixes across agents, networking, and UI.
Official source → - FEATURE
Claude Sonnet 5 is now the default model in Claude Code
Claude Code 2.1.197 switches the default model to Claude Sonnet 5, which ships with a native 1M-token context window and promotional pricing of $2/$10 per million tokens through August 31.
Official source → - FEATURE
Claude Code 2.1.197: Claude Sonnet 5 is now the default model
Claude Code 2.1.197 switches the default model to Claude Sonnet 5, which brings a native 1M-token context window and promotional pricing of $2/$10 per million tokens through August 31.
Official source → - FEATURE
Claude Code 2.1.196
Version 2.1.196 of Claude Code adds org-managed default models, readable session names, and clickable file attachments, while shipping a broad round of security hardening, reliability improvements, and bug fixes across agents, background sessions, MCP, and the terminal UI.
Official source → - FEATURE
Claude Code 2.1.160: Safety prompts, WSL clipboard fix, and major background-agent overhaul
Version 2.1.160 adds confirmation prompts before Claude writes to shell startup files or build-tool configs that could trigger unintended code execution, and ships a broad wave of bug fixes targeting background agents, Windows/WSL behavior, and terminal compatibility.
Official source → - FEATURE
Claude Code 2.1.163: Version Gates, Plugin List, Hook Feedback & 18 Fixes
Claude Code 2.1.163 adds managed version-range enforcement, a /plugin list command, a clipboard shortcut in /btw, richer hook feedback via additionalContext, and resolves 18 bugs spanning Windows, Bedrock/Vertex, Bazel, and agent session handling.
Official source → - FEATURE
Claude Code 2.1.195: Mouse control, hook fixes, voice dictation improvements
Version 2.1.195 ships a new environment variable to disable mouse interactions in fullscreen mode, fixes hook substring-matching bugs and several voice dictation issues, and improves background agent reliability and the agents list UI.
Official source → - FEATURE
Claude Code 2.1.193
Version 2.1.193 of Claude Code adds finer shell-command routing controls, richer auto-mode denial visibility, live bash path autocomplete, and OpenTelemetry response logging — plus a wave of background-agent reliability fixes.
Official source → - FEATURE
Claude Code 2.1.191: /rewind, CPU savings, MCP improvements, and many fixes
Version 2.1.191 adds /rewind support to resume conversations from before a /clear, cuts streaming CPU usage by ~37%, and ships a wave of bug fixes and reliability improvements across agents, MCP, hooks, and the terminal UI.
Official source → - FEATURE
Claude Code 2.1.187: Credential Sandboxing, Org Model Restrictions & Bug Fixes
Version 2.1.187 of Claude Code adds credential-file sandboxing, org-level model restrictions in the model picker, and mouse-click support in fullscreen menus, alongside roughly 17 bug fixes covering MCP timeouts, structured output loops, CJK text encoding, and more.
Official source → - FEATURE
Claude Code 2.1.186
Version 2.1.186 of Claude Code adds CLI-based MCP authentication, bash-command auto-response, and a Skills section in the plugin tab, alongside more than two dozen bug fixes and workflow improvements.
Official source → - FEATURE
Claude Code 2.1.185: Clearer stream-stall messaging and longer wait threshold
The stream-stall hint now shows a friendlier 'Waiting for API response · will retry in …' message instead of the previous 'No response from API · Retrying in …' wording, and the hint now triggers after 20 seconds of silence rather than 10.
Official source → - FEATURE
Claude Code 2.1.166: Fallback Models, Glob Deny Rules & Hardened Security
Claude Code 2.1.166 adds configurable fallback models (up to three, tried in order when the primary is overloaded), glob pattern support in deny rules, and hardens cross-session messaging so relayed messages can no longer carry user authority.
Official source → - FEATURE
Claude Code 2.1.183
Claude Code 2.1.183 adds auto-mode safety blocks for destructive git and infrastructure commands, a new attribution setting to omit session links from commits and PRs, and a /config --help command, alongside a batch of bug fixes covering subagents, TUI rendering, tmux, and MCP authentication.
Official source → - FEATURE
Claude Code 2.1.173: Fable 5 model name fix + Windows sandbox warning fix
Version 2.1.173 of Claude Code strips the `[1m]` suffix from Fable 5 model names automatically, since 1M context is included by default, and fixes a false 'sandbox dependencies missing' warning that appeared on Windows at startup.
Official source → - FEATURE
Claude Code 2.1.169: Safe Mode, /cd Command, Post-Session Hooks & Fixes
Claude Code 2.1.169 adds a --safe-mode flag for clean-slate troubleshooting, a /cd command to change working directory mid-session without breaking prompt cache, and a post-session lifecycle hook for self-hosted runners. The release also ships roughly two dozen bug fixes covering MCP policy enforcement, Windows performance regressions, background agent reliability, and more.
Official source → - FEATURE
Claude Code 2.1.181: Config Shortcuts, Smarter Retries & Dozens of Fixes
Claude Code 2.1.181 adds a `/config key=value` prompt command, a new `CLAUDE_CLIENT_PRESENCE_FILE` env var for suppressing mobile push notifications, Apple Events sandbox opt-in on macOS, and upgrades the bundled Bun runtime to 1.4. The release also ships improved streaming, subagent panel polish, and fixes over 30 bugs ranging from startup delays to truncated file writes.
Official source → - FEATURE
Claude Code 2.1.179: Connection, Scrolling, and UI Bug Fixes
Version 2.1.179 of Claude Code ships nine targeted fixes covering mid-stream connection drops, WSL2 mouse-wheel scrolling, an oversized Bash tool description on Linux, and several subagent and UI reliability issues.
Official source → - FEATURE
Claude Code 2.1.178: Smarter Permissions, Nested Skills & Bug Fixes
Claude Code 2.1.178 adds parameter-based permission rules, improved nested skills and directory resolution, safer subagent auto-mode evaluation, and fixes more than a dozen crashes and edge-case bugs.
Official source → - FEATURE
Claude Code 2.1.174: Model picker fixes, GovCloud routing, and VSCode usage stats
Version 2.1.174 of Claude Code ships a new scroll-acceleration setting, fixes the /model picker to correctly show which model Default resolves to on each plan, and adds per-feature usage attribution in the VSCode account dialog. A batch of infrastructure and reliability bugs are also resolved.
Official source → - FEATURE
Claude Code 2.1.172
Claude Code 2.1.172 ships nested sub-agent spawning up to 5 levels deep, a plugin search bar, and a broad wave of bug fixes covering context-limit sessions, agent panel state, model pickers, wildcard permissions, and performance improvements.
Official source → - FEATURE
Claude Code Auto Mode Now Available on Bedrock, Vertex, and Foundry
Claude Code 2.1.158 brings Auto mode to Bedrock, Vertex, and Foundry for Opus 4.7 and Opus 4.8. Enable it by setting the environment variable `CLAUDE_CODE_ENABLE_AUTO_MODE=1`.
Official source → - FEATURE
Claude Code 2.1.176: Session Language, Footer Links, Bedrock Caching & Fixes
Claude Code 2.1.176 adds multilingual session titles, a new footerLinksRegexes setting, smarter Bedrock credential caching, and closes out a large batch of bugs across Remote Control, background sessions, hooks, clipboard, and Windows.
Official source → - FEATURE
Claude Code 2.1.175: enforceAvailableModels Managed Setting
Version 2.1.175 adds the `enforceAvailableModels` managed setting, which locks the Default model to the administrator-defined allowlist and prevents users or projects from overriding it.
Official source → - FEATURE
GET /v1/environments/{id}/work now available on Claude Platform on AWS
The endpoint that lists pending work for a self-hosted sandbox is now available on Claude Platform on AWS. Access requires the `GetEnvironment` IAM action.
Official source → - FEATURE
Claude Code: Background agents now preserve current permission mode
In Claude Code 2.1.141, background agents launched via `/bg` or `←←` no longer revert to the default permission mode — they inherit whatever permission mode was active when you launched them.
Official source → - FEATURE
Claude Code: Fast mode on Opus 4.8 now cheaper — 2x rate, 2.5x speed
Claude Code 2.1.154 brings a significant price drop for Fast mode on Opus 4.8: it now runs at 2x the standard rate while delivering 2.5x the speed, down from its previous higher cost.
Official source → - FEATURE
Claude Code: Background agent sessions now respect permissions.defaultMode
In Claude Code 2.1.143, background sessions launched via `claude agents` now correctly honor the `permissions.defaultMode` setting from settings.json. Previously, these sessions were hard-coded to auto mode regardless of your configuration.
Official source → - MODEL
Claude Fable 5: 1M Context, Adaptive Thinking, Mythos-Class Power
Claude Fable 5 (claude-fable-5) is Anthropic’s most capable widely released model — a Mythos-class model made safe for general use, with capabilities Anthropic says exceed any model it has ever made generally available. It ships with a 1M-token context window by default, 128K max output tokens, and always-on adaptive thinking, priced at $10/$50 per million tokens (input/output). Generally available June 9, 2026 on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry; Claude Code users get it by updating to v2.1.170.
Official source → - FEATURE
Claude Code 2.1.165: Bug Fixes and Reliability Improvements
Claude Code version 2.1.165 ships bug fixes and reliability improvements. No specific changes were detailed in the release notes.
Official source → - FEATURE
Claude Code 2.1.161: Metrics slicing, parallel tool fixes, Linux clipboard
Version 2.1.161 of Claude Code ships OTEL resource-attribute labels for custom metric dimensions, independent parallel tool-call results so a failed Bash command no longer kills its batch-mates, and improved Linux clipboard support in fullscreen mode, alongside a large round of bug fixes.
Official source → - FEATURE
Claude Code 2.1.162: Agents view fixes, tool improvements, startup polish
Version 2.1.162 brings a wave of bug fixes and UX refinements across the Claude Code CLI, including improved `claude agents` reliability, better permission rule handling, and a quieter, clearer startup experience.
Official source → Full refresh
All feature guides re-validated against the latest Claude documentation.
Full refresh
All feature guides re-validated against the latest Claude documentation.
Claude Master launches
First public release — all 85 features researched, written, and cross-validated against official Anthropic documentation.
- FEATURE
Managed Agents webhooks, multiagent orchestration & self-hosted sandboxes on AWS
Claude Managed Agents webhooks, multiagent orchestration, and self-hosted sandboxes are now available on Claude Platform on AWS. New IAM actions and the `AnthropicSelfHostedEnvironmentAccess` managed policy support the rollout.
Official source → - FEATURE
Ultracode in Claude Code: Fan a Big Task Across Many Agents
Claude Code now has "ultracode," a session setting that turns reasoning up to its highest setting (xhigh) and lets Claude decide when a task is big enough to split across many cooperating sub-agents instead of working step by step. You can turn it on for a whole session with /effort ultracode, or trigger it for a single task by putting the word "ultracode" in your prompt. It's built for jobs like auditing every file in a codebase, large migrations, or research that needs sources cross-checked, and you can turn it off anytime in /config.
Official source → - MODEL
Claude Opus 4.8: 1M context, adaptive thinking, mid-conversation system messages
Claude Opus 4.8 launches as Anthropic's most capable generally available model, adding a 1M token context window by default, adaptive thinking to reduce wasted reasoning tokens, and mid-conversation system messages that preserve prompt cache hits when instructions change mid-session.
Official source →