ClaudeCodeChanges

New version Docs only
S M T W T F S
1 2 3 4 5 6 7
8 9 10 11 12 13 14
15 16 17 18 19 20 21
22 23 24 25 26 27 28
29 30 31

Claude docs changes for March 4th, 2026 [diff]

Executive Summary

  • Claude Code 2.1.68 expands effort level support to Sonnet 4.6 (previously Opus 4.6 only), re-introduces the ultrathink keyword for high-effort mode, and removes Opus 4/4.1 from the first-party API
  • The Agent SDK permissions documentation was significantly expanded with a new dontAsk mode (TypeScript only), clearer distinction between allow/deny rule evaluation, and guidance on how bypassPermissions interacts with allowed_tools
  • Dozens of API documentation pages had their primary code examples switched from Python to Shell (curl), making them more accessible to users of any language
  • Multiple new MCP server integrations were added to both the Claude Code and API documentation (Gusto, LSEG, Mem, Pylon, Daloopa, Similarweb, Bigdata.com, and more)

New Claude Code versions

2.1.68

Existing feature improvements

  • Opus 4.6 now defaults to medium effort for Max and Team subscribers; medium effort is the sweet spot between speed and thoroughness and can be changed with /model
  • Effort level support expanded to Sonnet 4.6 (in addition to Opus 4.6)
  • Re-introduced the ultrathink keyword to enable high effort for the next turn

Major bug fixes

  • Removed Opus 4 and Opus 4.1 from Claude Code on the first-party API; users with these models pinned are automatically moved to Opus 4.6

Claude Code changes

Changed documents

common-workflows [Source]

  • Updated effort level description to note that both Opus 4.6 and Sonnet 4.6 support effort control (previously only Opus 4.6 was mentioned) [line 845] [Source]

mcp [Source]

  • Added new MCP server: Smartsheet — Analyze and manage Smartsheet data with Claude
  • Added new MCP server: Windsor.ai — Connect 325+ marketing, analytics and CRM data sources
  • Added new MCP server: Mem — The AI notebook for everything on your mind
  • Various MCP server entries were reordered

model-config [Source]

  • Effort level support now documented for both Opus 4.6 and Sonnet 4.6; removed note saying effort was only supported on Opus 4.6 [line 145] [Source]

settings [Source]

  • CLAUDE_CODE_EFFORT_LEVEL environment variable description updated to reflect support for both Opus 4.6 and Sonnet 4.6 [line 945] [Source]

API changes

Changed documents

content-moderation [Source]

  • Removed redundant client initialization boilerplate from multiple code examples, cleaning up the guide

legal-summarization [Source]

  • Removed redundant client initialization boilerplate from code examples

migration-guide [Source]

  • All Python migration code examples replaced with Shell (curl) equivalents, covering low-effort, medium-effort with extended thinking, and adaptive thinking scenarios [lines 232-362] [Source]

ticket-routing [Source]

  • Removed redundant client initialization boilerplate from code examples

overview (agent-sdk) [Source]

  • Minor cleanup

permissions [Source]

  • Significantly restructured permission evaluation flow: deny rules are now evaluated separately before allow rules, and the ordering is now: hooks → deny rules → permission mode → allow rules → canUseTool [lines 17-44] [Source]
  • New section added explaining allowed_tools and disallowed_tools behavior with a table and practical examples [lines 46-72] [Source]
  • New dontAsk permission mode documented (TypeScript only): converts permission prompts into denials for fixed, explicit tool surfaces [lines 80-84] [Source]
  • Clarification that allowed_tools does NOT constrain bypassPermissions — all tools are still approved unless explicitly in disallowed_tools [lines 66-68] [Source]
  • New dedicated subsection for dontAsk mode with guidance on Python equivalent workarounds [lines 131-140] [Source]

python (agent-sdk) [Source]

  • Minor cleanup

quickstart (agent-sdk) [Source]

  • Minor cleanup

skills (agent-sdk) [Source]

  • Minor cleanup

typescript (agent-sdk) [Source]

  • Minor cleanup

quickstart (agent-skills) [Source]

  • Minor cleanup

mcp-connector [Source]

  • Minor cleanup

remote-mcp-servers [Source]

  • Added new MCP server: Gusto — Query and analyze your Gusto data (https://mcp.api.gusto.com/anthropic)
  • Added new MCP server: LSEG — Access best in class data & analytics across a broad spectrum of asset classes (https://api.analytics.lseg.com/lfa/mcp/server-cl)
  • Added new MCP server: Bigdata.com — Access real-time financial data (https://mcp.bigdata.com/)
  • Added new MCP server: Daloopa — Financial fundamental data and KPIs with hyperlinks (https://mcp.daloopa.com/server/mcp)
  • Added new MCP server: Pylon — Search and manage Pylon support issues (https://mcp.usepylon.com/)
  • Added new MCP server: Similarweb — Real time web, mobile app, and market data (https://mcp.similarweb.com)
  • Added new MCP server: WordPress.com — Secure AI access to manage your WordPress.com sites (https://public-api.wordpress.com/wpcom/v2/mcp/v1)
  • Added new MCP server: Yardi Virtuoso — Real-time Yardi data & insights (https://mcp.virtuoso.ai/mcp)
  • Various existing server entries were reordered

code-execution-tool [Source]

  • Container reuse example replaced with Shell (curl) equivalent, showing how to extract container ID with jq and pass it to a second request [lines 506-553] [Source]

computer-use-tool [Source]

implement-tool-use [Source]

  • Minor cleanup

overview (tool-use) [Source]

  • Minor cleanup

programmatic-tool-calling [Source]

  • Minor cleanup

tool-search-tool [Source]

  • Fixed cache_control structure in multi-turn conversation example: moved cache_control inside the content block array (correct API structure) [lines 403-414] [Source]
  • Fixed usage.get() call to use proper SDK attribute access: response2.usage.cache_read_input_tokens or 0

web-fetch-tool [Source]

  • Fixed cache_control structure in multi-turn example: moved inside content block array (matching correct API structure) [lines 379-390] [Source]
  • Fixed usage.get() call to use proper SDK attribute access

web-search-tool [Source]

  • Fixed cache_control structure in multi-turn example: moved inside content block array (correct API structure) [lines 324-336] [Source]
  • Fixed usage.get() call to use proper SDK attribute access

client-sdks [Source]

  • Language tab ordering updated: C# now appears before Java, Ruby moved after PHP

openai-sdk [Source]

  • API key now read from environment variable os.environ.get("ANTHROPIC_API_KEY") instead of a literal placeholder string [line 30] [Source]
  • Extended thinking example now includes a concrete messages parameter instead of ... [line 76] [Source]

python (sdk) [Source]

  • Removed redundant client initialization boilerplate from multiple code snippets throughout the guide
  • Added a proper example instantiation before the if response.my_field is None null-distinction example [line 537] [Source]

typescript (sdk) [Source]

  • Removed redundant import Anthropic and client initialization boilerplate from many code snippets
  • Fixed Message Batches examples to use client instead of anthropic variable name for consistency [lines 272-302] [Source]
  • fetchAllMessageBatches function now uses typed parameter params: Record<string, unknown> instead of untyped params [line 491] [Source]

adaptive-thinking [Source]

  • Adaptive thinking + effort level example replaced with Shell (curl) equivalent [lines 63-92] [Source]

batch-processing [Source]

  • Polling loop example replaced with Shell (curl/bash) equivalent using a until loop [lines 169-197] [Source]
  • List batches example replaced with Shell (curl) equivalent with manual pagination loop using jq [lines 205-244] [Source]
  • Cancel batch example replaced with Shell (curl) equivalent [lines 303-316] [Source]

citations [Source]

  • Minor cleanup

claude-in-microsoft-foundry [Source]

  • API key authentication example replaced with Shell (curl) equivalent [lines 98-118] [Source]
  • Entra ID (Azure AD) authentication example replaced with Shell equivalent using az account get-access-token [lines 126-150] [Source]
  • Language tab ordering updated: C# now appears before Java

claude-on-amazon-bedrock [Source]

  • Language tab ordering updated: C# now appears before Java

claude-on-vertex-ai [Source]

  • Basic text generation example replaced with Shell (curl) equivalent using gcloud auth print-access-token [lines 69-94] [Source]
  • Language tab ordering updated: Go now appears before Java

compaction [Source]

  • Added a complete example request showing how to trigger manual compaction using context_management with compact_20260112 edit type [lines 234-243] [Source]

context-editing [Source]

  • Minor cleanup

effort [Source]

extended-thinking [Source]

  • Minor cleanup

fast-mode [Source]

  • Minor cleanup

files [Source]

  • Minor cleanup

handling-stop-reasons [Source]

  • Added full client initialization and message creation context to the end_turn stop reason code example [lines 39-47] [Source]
  • Added full setup code (client, tool definition, execute_tool function) to the tool_use stop reason example [lines 180-205] [Source]
  • Added complete client and request setup to the error handling and streaming examples [lines 387-400] [Source]

multilingual-support [Source]

  • Minor cleanup

prompting-tools [Source]

  • Minor cleanup

search-results [Source]

  • Direct search results example replaced with Shell (curl) equivalent, showing the full JSON structure for search_result content blocks [lines 187-244] [Source]

skills-guide [Source]

  • All major code examples converted from Python to Shell (curl): basic skill usage, file download, multi-turn pause_turn handling, combining multiple skills, skill upload, listing/retrieving/deleting skills, versioning, and prompt caching examples [throughout]
  • Error handling section retains Python as it uses the SDK's exception model (no shell equivalent)

streaming [Source]

  • Minor cleanup

structured-outputs [Source]

  • Language tab ordering updated: C# now appears before Java, Ruby moved after PHP

token-counting [Source]

  • Basic token counting example replaced with Shell (curl) equivalent [lines 26-46] [Source]
  • Token counting with tools example replaced with Shell (curl) equivalent [lines 53-90] [Source]

overview (resources) [Source]

  • Minor cleanup