Claude docs changes for March 12th, 2026 [diff]

Executive Summary

  • New modelOverrides setting (2.1.73) lets enterprise admins map individual model versions to provider-specific IDs such as Bedrock inference profile ARNs, enabling per-version routing and governance
  • New autoMemoryDirectory setting (2.1.74) allows configuring a custom directory for auto-memory storage; /context command gains actionable optimization tips
  • API Tier 4 monthly spend threshold raised dramatically from $5,000 to $200,000
  • /output-style command deprecated in favor of /config; output style is now locked at session start to enable prompt caching
  • 1-hour prompt caching now supported on Google Cloud's Vertex AI (previously only Claude API and Microsoft Foundry)

New Claude Code versions

2.1.73

New features

  • Added modelOverrides setting to map model picker entries to custom provider model IDs (e.g., Bedrock inference profile ARNs), enabling per-version routing for enterprise environments
  • Added actionable guidance when OAuth login or connectivity checks fail due to SSL certificate errors (corporate proxies, NODE_EXTRA_CA_CERTS)

Existing feature improvements

  • Changed default Opus model on Bedrock, Vertex, and Microsoft Foundry to Opus 4.6 (was Opus 4.1)
  • Deprecated /output-style command — use /config instead; output style is now fixed at session start for better prompt caching
  • Improved Up arrow after interrupting Claude — now restores the interrupted prompt and rewinds the conversation in one step
  • Improved /effort to work while Claude is responding, matching /model behavior
  • Improved IDE detection speed at startup and clipboard image pasting performance on macOS
  • Improved voice mode to automatically retry transient connection failures during rapid push-to-talk re-press

Major bug fixes

  • Fixed freezes and 100% CPU loops triggered by permission prompts for complex bash commands
  • Fixed a deadlock that could freeze Claude Code when many skill files changed at once (e.g., during git pull in a repo with a large .claude/skills/ directory)
  • Fixed Bash tool output being lost when running multiple Claude Code sessions in the same project directory
  • Fixed subagents with model: opus/sonnet/haiku being silently downgraded to older model versions on Bedrock, Vertex, and Microsoft Foundry
  • Fixed SessionStart hooks firing twice when resuming a session via --resume or --continue
  • Fixed JSON-output hooks injecting no-op system-reminder messages into the model's context on every turn
  • Fixed Linux sandbox failing to start with "ripgrep (rg) not found" on native builds
  • Fixed Linux native modules not loading on Amazon Linux 2 and other glibc 2.26 systems
  • VSCode: Fixed HTTP 400 errors for users behind proxies or on Bedrock/Vertex with Claude 4.5 models

2.1.74

New features

  • Added actionable suggestions to /context command — identifies context-heavy tools, memory bloat, and capacity warnings with specific optimization tips
  • Added autoMemoryDirectory setting to configure a custom directory for auto-memory storage

Existing feature improvements

  • Changed --plugin-dir so local dev copies now override installed marketplace plugins with the same name (unless force-enabled by managed settings)

Major bug fixes

  • Fixed memory leak where streaming API response buffers were not released when the generator was terminated early, causing unbounded RSS growth on the Node.js/npm code path
  • Fixed managed policy ask rules being bypassed by user allow rules or skill allowed-tools
  • Fixed full model IDs (e.g., claude-opus-4-5) being silently ignored in agent frontmatter model: field and --agents JSON config
  • Fixed MCP OAuth authentication hanging when the callback port is already in use
  • Fixed MCP OAuth refresh never prompting for re-auth after the refresh token expires for OAuth servers that return errors with HTTP 200 (e.g., Slack)
  • Fixed SessionEnd hooks being killed after 1.5s on exit regardless of hook.timeout — now configurable via CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS
  • Fixed Hebrew, Arabic, and other RTL text not rendering correctly in Windows Terminal, conhost, and VS Code integrated terminal
  • Fixed LSP servers not working on Windows due to malformed file URIs
  • VSCode: Fixed delete button not working for Untitled sessions

Claude Code changes

Changed documents

amazon-bedrock [Source]

  • New section added explaining how to use modelOverrides to map multiple model versions within the same family to distinct Bedrock inference profile ARNs. [lines 209-230] [Source]

interactive-mode [Source]

  • Updated /config command description to clarify it adjusts theme, model, output style, and other preferences. [line 85] [Source]
  • Removed /output-style from the slash commands table (deprecated in favor of /config).

model-config [Source]

  • New section "Override model IDs per version" documenting the modelOverrides setting, which maps individual Anthropic model IDs to provider-specific strings for per-version routing on Bedrock, Vertex, and Foundry. [lines 227-252] [Source]

output-styles [Source]

  • Instructions for changing output style updated: now uses /config menu instead of the deprecated /output-style command. [lines 35-58] [Source]
  • Added note that output style is fixed at session start so the system prompt stays stable, enabling prompt caching to reduce latency and cost. [lines 55-58] [Source]

settings [Source]

  • Added modelOverrides to the settings reference table, describing how to map Anthropic model IDs to provider-specific model IDs such as Bedrock inference profile ARNs. [line 158] [Source]

API changes

Changed documents

build-with-claude/overview [Source]

  • Google Cloud's Vertex AI added to the list of platforms supporting 1-hour prompt caching (previously only Claude API and Microsoft Foundry). [line 75] [Source]

api/rate-limits [Source]

  • Tier 4 monthly spend threshold increased from $5,000 to $200,000. [line 38] [Source]

test-and-evaluate/strengthen-guardrails/handle-streaming-refusals [Source]

  • Removed the special test string (ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL_...) that could be used to trigger refusals in testing.