Claude Code Desktop gains major new features: live app preview with embedded browser, GitHub PR monitoring with auto-fix and auto-merge, and a new "Review code" button in diff view
Fast mode pricing simplified to a single flat rate ($30/$150 per MTok) with full 1M context window included — no more premium for requests over 200K tokens
Version 2.1.50 brings numerous memory leak fixes for long sessions, new WorktreeCreate/WorktreeRemove hook events, and agent worktree isolation support
Programmatic tool calling updated to new tool version code_execution_20260120; web search and web fetch tools can now be called programmatically
New MCP servers added across both Claude Code and API docs, including Granola, monday.com, Asana, PubMed, Box, Miro, Microsoft Learn, and many more
Added new "Preview your app" section documenting the live dev server preview feature, including how Claude auto-starts servers, takes screenshots, inspects the DOM, and iterates on findings [line 67] [Source]
Added full "Configure preview servers" reference with .claude/launch.json format, autoVerify option, all configuration fields, and examples for Next.js, monorepos, and Node.js scripts [line 147] [Source]
Added "Review your code" feature: clicking "Review code" in the diff view toolbar has Claude evaluate diffs and leave inline comments [line 84] [Source]
Added "Monitor pull request status" section: CI status bar with auto-fix and auto-merge toggles powered by the GitHub CLI [line 106] [Source]
Windows ARM64 devices are now fully supported (previous note about local session limitations removed) [line 504] [Source]
Cowork tab is now available on Windows on all supported hardware (previously macOS Apple Silicon only) [line 507] [Source]
Managed settings key updated from permissions.disableBypassPermissionsMode to disableBypassPermissionsMode; added reference to allowManagedPermissionRulesOnly and allowManagedHooksOnly [line 351] [Source]
Updated note on MAX_THINKING_TOKENS: on Opus, setting is ignored except for 0 because adaptive reasoning controls thinking depth [line 322] [Source]
Added note about multi-environment behavior: when using both the bash tool and the code execution tool, state is not shared between them, with a link to guidance [line 262] [Source]
Added new "Using code execution with other execution tools" section with system prompt guidance for multi-computer environments where Claude may confuse separate execution contexts [line 84] [Source]
Web search and web fetch tools removed from the "not supported for programmatic calling" list — they are now supported
Page renamed from "Create strong empirical evaluations" to "Define success criteria and build evaluations"; merged in the content previously on the separate "Define success criteria" page [line 1] [Source]
Added detailed guidance on writing good success criteria (Specific, Measurable, Achievable, Relevant) with examples [line 14] [Source]
Fast mode is now ZDR (Zero Data Retention) eligible [line 8] [Source]
Pricing simplified: removed tiered pricing for >200K tokens; now a single flat rate of $30/MTok input and $150/MTok output for the full context window [line 48] [Source]
Added new "Keep Claude in character" section with guidance on using system prompts for role-based applications and an enterprise chatbot example [line 43] [Source]
Replaced the simple "Get started" links section with a numbered 4-step "Recommended path for new developers" (quickstart → Messages API → models → features) [line 14] [Source]
Fast mode pricing simplified: removed the two-tier table (≤200K and >200K tokens); now a single $30/MTok input, $150/MTok output rate that applies to the full context window [line 72] [Source]
Clarified that standard-speed long context pricing does not apply to fast mode requests [line 110] [Source]
Updated tool version identifier from code_execution_20250825 to code_execution_20260120 throughout all examples and reference tables [line 14] [Source]