Claude Code Default Effort Raised to High

Default Effort Level Raised to High

With Claude Code v2.1.94, Anthropic has changed the default thinking effort level from medium to high for the following user types:

• API-key users
• Amazon Bedrock, Vertex AI, and Microsoft Foundry users
• Team plan subscribers
• Enterprise plan subscribers

Why This Matters

Effort level controls how deeply Claude reasons before responding. At high effort, Claude generates extended thinking blocks — internal reasoning traces that allow it to analyze multi-file architectures, plan complex refactors, and reason about downstream consequences before writing a single line of code.

The previous default of "medium" was introduced in early 2026 as a cost-and-speed optimization. The decision proved controversial: a significant portion of the developer community reported degraded output quality on non-trivial tasks, and the change generated sustained criticism on X (Twitter), Hacker News, and GitHub Issues. Wolfram Ravenwolf, a prominent AI developer, called it "not a good change" and noted that the downgrade wasn't even persisted — selecting high effort manually would revert to medium on restart.

With v2.1.94, Anthropic has reversed course. For the affected user types, every new session now defaults to high effort without any manual configuration needed.

Controlling Effort Level

Users who prefer a different setting can adjust it at any time:

• Per session: use the /effort command and select low, medium, high, or max
• Persistent override: set CLAUDE_CODE_EFFORT_LEVEL in settings.json under the env key
• One-time max reasoning: type ultrathink in any prompt to temporarily engage maximum effort for that turn

Note that Max plan subscribers retain their own separate default and are not affected by this change.

Additional Improvements in v2.1.94

Rate-Limit Recovery

Agents no longer appear permanently stuck after receiving a 429 rate-limit response with a long Retry-After header. Previously, the tool would silently wait for the full retry delay, leaving sessions that appeared frozen. The error now surfaces immediately, giving users visibility and control.

Terminal and Rendering Fixes

Several long-standing rendering issues were resolved:

• Scrollback no longer shows the same diff repeated in long-running sessions
• Multi-line user prompts no longer incorrectly indent wrapped lines under the ❯ caret
• Shift+Space now inserts a real space instead of the literal word "space" in search inputs
• Clicking hyperlinks inside tmux running in VS Code, Hyper, or Tabby terminals no longer opens two browser tabs
• The FORCE_HYPERLINK environment variable is now respected when set via settings.json

Cross-Worktree Resume

--resume can now resume sessions from other worktrees of the same repository directly, instead of printing a cd command and requiring the user to navigate manually first.