Claude docs changes for February 27th, 2026 [diff]

Executive Summary

  • Remote Control is now restricted to Max plans only (Pro support coming soon), narrowing availability compared to the previous Pro+Max rollout
  • The & prefix for sending tasks to the web has been replaced by the --remote flag, a breaking workflow change for terminal-to-web handoffs
  • New CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 and CLAUDE_CODE_DISABLE_FAST_MODE=1 environment variables give administrators control over adaptive reasoning and fast mode
  • Agent SDK hooks expanded: PostToolUseFailure, SubagentStart, PermissionRequest, and Notification are now available in the Python SDK, plus new TypeScript-only hooks (Setup, TeammateIdle, TaskCompleted, ConfigChange, WorktreeCreate, WorktreeRemove)
  • Claude Sonnet 4.6 now has its own prompt caching minimum token threshold of 2048 (separate from other Sonnet models)

New Claude Code versions

2.1.61

Major bug fixes

  • Fixed concurrent writes corrupting config file on Windows

2.1.62

Major bug fixes

  • Fixed prompt suggestion cache regression that reduced cache hit rates

Claude Code changes

Changed documents

Best practices [Source]

  • Terminology changed from "headless mode" to "non-interactive mode" throughout. [line 498] [Source]
  • Permissions link corrected from settings.md to permissions.md. [line 233] [Source]

Claude Code on the web [Source]

  • The & message prefix for sending tasks to the web has been replaced by --remote flag (e.g., claude --remote "Fix the bug"). [line 76] [Source]
  • The /remote-env command now only references --remote (the & prefix is no longer mentioned). [line 233] [Source]

Common workflows [Source]

  • MAX_THINKING_TOKENS is now ignored on both Opus 4.6 and Sonnet 4.6 (previously only Opus 4.6). [line 857] [Source]
  • New CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 environment variable documented for disabling adaptive thinking and reverting to a fixed thinking budget. [line 858] [Source]

Discover plugins [Source]

  • Added in-app submission forms for the official Anthropic plugin marketplace: claude.ai/settings/plugins/submit and platform.claude.com/plugins/submit. [line 39] [Source]

Fast mode [Source]

  • New CLAUDE_CODE_DISABLE_FAST_MODE=1 environment variable added as an option to disable fast mode entirely. [line 90] [Source]

Interactive mode [Source]

  • /rename [name] command updated: when called without a name argument, it now auto-generates a name from conversation history (requires at least one message). [line 94] [Source]
  • Clarification that input history resets on /clear but the previous session's conversation is preserved and can be resumed. [line 188] [Source]

MCP [Source]

  • New MCP servers added to the marketplace: Airwallex Developer, Outreach, and several reorderings. [line 669] [Source]

Model config [Source]

  • Added documentation for CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 to disable adaptive reasoning on Opus 4.6 and Sonnet 4.6, reverting to fixed thinking budget. [line 152] [Source]

Plugins [Source]

  • New section added for submitting plugins to the official Anthropic marketplace via in-app forms on claude.ai and platform.claude.com. [line 391] [Source]

Remote control [Source]

  • Remote Control availability narrowed: now only available on Max plans (Pro plan support is coming soon, previously it was available on both Pro and Max). [line 1] [Source]

Settings [Source]

  • New CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING environment variable: set to 1 to disable adaptive reasoning for Opus 4.6 and Sonnet 4.6, falling back to fixed MAX_THINKING_TOKENS budget. [line 910] [Source]
  • New CLAUDE_CODE_DISABLE_FAST_MODE environment variable: set to 1 to disable fast mode. [line 913] [Source]

API changes

Changed documents

Agent SDK - Cost tracking [Source]

  • Major rewrite clarifying Python vs TypeScript SDK differences for cost tracking. Python only provides accumulated totals on the result message; TypeScript provides per-step breakdowns. [line 1] [Source]
  • Added a diagram illustrating the message stream from a query() call with token usage at each step.
  • Python SDK reference added alongside the existing TypeScript SDK reference.

Agent SDK - Hooks [Source]

  • PostToolUseFailure, SubagentStart, PermissionRequest, and Notification hooks are now available in the Python SDK (previously TypeScript-only). [line 84] [Source]
  • New TypeScript-only hooks added: Setup, TeammateIdle, TaskCompleted, ConfigChange, WorktreeCreate, WorktreeRemove. [line 95] [Source]
  • The introductory section was rewritten to explain hooks as a step-by-step flow (event fires → SDK collects hooks → matchers filter → callback executes → decision returned).

Agent SDK - TypeScript [Source]

  • SDKAssistantMessageError no longer includes 'max_output_tokens' as a valid error value. [line 585] [Source]

Agents and tools - Remote MCP servers [Source]

  • New MCP servers added: Airtable, Common Room, LSEG, Jotform, Airwallex Developer, Supermetrics, MSCI. [line 862] [Source]

Build with Claude - Prompt caching [Source]

  • Claude Sonnet 4.6 now has its own minimum cacheable prompt length of 2048 tokens, separate from other Sonnet models (which remain at 1024). [line 256] [Source]