Claude Code release notes: 2.1.267 caps reasoning effort

Claude Code 2.1.267 adds maxEffortLevel, a setting that caps the model's reasoning effort. Users can still choose a lower level. The cap also works with Amazon Bedrock, Google Vertex, and Microsoft Foundry.

The version was published on September 9 and was the latest on npm when checked on September 10.

Sources: Claude Code 2.1.267 release notes. Claude Code in the npm registry.

How the cap works in Claude Code 2.1.267

Effort is the level that controls how much the model reasons. Claude Code applies the cap before each request. If a user chooses a higher level through /effort, the model picker, or --effort, the request runs at the cap instead. This also covers levels from the CLAUDE_CODE_EFFORT_LEVEL environment variable and the model's default.

The setting accepts low, medium, high, xhigh, and max. The value max means no cap. Leaving the setting unset also imposes no cap.

Source: Anthropic's maxEffortLevel reference.

Set a limit in the settings file

Human step: With Claude Code 2.1.267 or later, add the following key to your existing ~/.claude/settings.json. Keep your other settings:

\{
  "maxEffortLevel": "medium"
\}

This caps effort at medium across your projects on that machine. For a shared project setting, use .claude/settings.json. Administrators can distribute the same key through the organization's managed settings.

Sources: Anthropic's settings reference. Settings files and scope.

Model exceptions and overlapping limits

You can also put maxEffortLevel under a specific model in modelSettings. The model's value replaces the top-level cap only within the same settings source. A per-model value of max therefore exempts the model from that source's cap, not from other organization limits.

When caps come from several settings files, the lowest applies. If the organization also has a separate effort limit for the model, the lower of the two limits applies.

Source: Anthropic's maxEffortLevel reference.