Claude docs changes for March 3rd, 2026 [diff]

Executive Summary

• Agent SDK stop-reasons doc overhauled: stop_reason on ResultMessage is now TypeScript-only; Python users need a stream-event workaround
• File checkpointing simplified: the CLAUDE_CODE_ENABLE_SDK_FILE_CHECKPOINTING environment variable is no longer required
• Data residency expanded: US-only inference now costs 1.1x standard rate for Opus 4.6+, with new workspace-level geo controls
• Fast mode constraints clarified: Opus 4.6 only, not available with Batch API or Priority Tier, and switching speeds invalidates prompt cache
• New MCP servers added: Customer.io, MailerLite, Clerk, Gusto, and Airtable

Claude Code changes

Changed documents

MCP [Source]

• Three new MCP servers added: Customer.io (explore customer data), MailerLite (email marketing assistant), and Clerk (authentication, organizations, and billing). [MCP servers list]
• PostHog changed from requiring a user-specific URL to providing a direct public endpoint (https://mcp.posthog.com/mcp). [line 501] [Source]

API changes

Changed documents

Adaptive Thinking [Source]

• Minor wording fix: "Claude will almost always think" corrected to "Claude almost always thinks" for accuracy.

Agent Skills - Best Practices [Source]

• Widespread punctuation normalization throughout the document (bold labels now use : consistently). No substantive content changes.

Agent Skills - Quickstart [Source]

• Section header renamed from "What are Agent Skills?" to "Agent Skills overview". Minor wording normalization throughout.

Batch Processing [Source]

• Results format description clarified: results are in .jsonl format (rewording from future to present tense).

Data Residency [Source]

• Added pricing details: US-only inference (inference_geo: "us") is priced at 1.1x the standard rate for Opus 4.6 and newer models across all token pricing categories. [line 82] [Source]
• Added distinction between inference geo (per-request, inference_geo parameter) and workspace geo (data at rest and endpoint processing, configured in Console). [lines 7-9] [Source]
• Documented two new workspace settings: allowed_inference_geos (restricts which geos a workspace can use) and default_inference_geo (fallback when inference_geo is omitted). [lines 63-64] [Source]
• Added note about Priority Tier: the 1.1x US inference multiplier also affects committed TPM burndown. [line 88] [Source]

Effort [Source]

• Added new guidance on effort level use cases: low effort is recommended for high-volume or latency-sensitive workloads and chat/non-coding tasks; high effort for tasks requiring maximum intelligence from Sonnet 4.6. [lines near effort table]
• Section header renamed from "When should I adjust the effort parameter?" to "When to adjust the effort parameter".

Extended Thinking [Source]

• Minor wording fix: clarified that using max effort on unsupported models "return an error" (present tense).

Fast Mode [Source]

• Added new Limitations section with multiple constraints: [lines 164-168] [Source]
• Fast mode is currently supported on Opus 4.6 only; sending speed: "fast" with another model returns an error
• Switching between fast and standard speed invalidates the prompt cache; requests at different speeds do not share cached prefixes
• Fast mode benefits focus on output tokens per second (OTPS), not time to first token (TTFT)
• Fast mode is not available with the Batch API
• Fast mode is not available with Priority Tier

File Checkpointing [Source]

• Removed requirement for the CLAUDE_CODE_ENABLE_SDK_FILE_CHECKPOINTING environment variable. File checkpointing now only requires setting enable_file_checkpointing=True (Python) or enableFileCheckpointing: true (TypeScript) in the SDK options. All code examples and the step-by-step guide updated accordingly. [line 57] [Source]
• Troubleshooting section updated: the "CheckpointNotFound" error cause now correctly references the missing enable_file_checkpointing flag rather than the old environment variable. [line 436] [Source]

Files [Source]

• Added a note that users can contact Anthropic for higher file storage limits.

Handle Stop Reasons [Source]

• Major rewrite: Clarified that stop_reason on ResultMessage is currently TypeScript-only. The Python SDK's ResultMessage does not include this field yet. [line 16] [Source]
• All primary code examples converted to TypeScript; a new "Read stop_reason in Python" section added with a workaround using include_partial_messages=True and scanning StreamEvent messages for message_delta events. [line 105] [Source]
• Added a new "Detect refusals" section with a TypeScript example for checking stop_reason === "refusal". [line near 70]

Implement Tool Use [Source]

• Added descriptive alt text to the tool_choice options diagram image. [line 483] [Source]

Migration Guide (Agent SDK) [Source]

• Corrected "Before" package version for @anthropic-ai/claude-code from ^1.0.0 to ^0.0.42. [line 56] [Source]
• Updated @anthropic-ai/claude-agent-sdk target version from ^0.1.0 to ^0.2.0. [line 66] [Source]
• Fixed Python "Before" migration example to correctly show from claude_code_sdk import (was incorrectly showing from claude_agent_sdk import). [line 130] [Source]

Remote MCP Servers [Source]

• New MCP servers added: Airtable (bring structured data to Claude), Gusto (query and analyze Gusto data), Customer.io (explore customer data), MailerLite (email marketing assistant), and Clerk (authentication, organizations, and billing).

SDK - C# [Source]

• Added integration notes for Bedrock (Anthropic.Bedrock) and Foundry (Anthropic.Foundry) client variants.

SDK - Go [Source]

• Added integration notes for Bedrock (github.com/anthropics/anthropic-sdk-go/bedrock) and Vertex AI (github.com/anthropics/anthropic-sdk-go/vertex) packages.

SDK - Java [Source]

• Added integration notes for Bedrock (com.anthropic:anthropic-java-bedrock), Vertex AI (com.anthropic:anthropic-java-vertex), and Foundry (com.anthropic:anthropic-java-foundry) packages.

SDK - Ruby [Source]

• Added integration notes for Bedrock (Anthropic::BedrockClient) and Vertex AI (Anthropic::VertexClient) clients.

SDK - TypeScript [Source]

• Added integration notes for Bedrock (@anthropic-ai/bedrock-sdk), Vertex AI (@anthropic-ai/vertex-sdk), and Foundry (@anthropic-ai/foundry-sdk) packages.

Skills Guide [Source]

• Added explicit limits: maximum 8 Skills per request and maximum 8MB total upload size per Skill. [lines 586-587] [Source]
• Documented that changing the Skills list in a container breaks the prompt cache.

Zero Data Retention [Source]

• Added 1M Token Context Window to the ZDR-eligible features table. The extended context window via anthropic-beta: context-1m-2025-08-07 is now confirmed ZDR-eligible.